diff --git a/.curlrc-homebrew-speech b/.curlrc-homebrew-speech
new file mode 100644
index 0000000..3d0452a
--- /dev/null
+++ b/.curlrc-homebrew-speech
@@ -0,0 +1 @@
+--http1.1
diff --git a/.dockerignore b/.dockerignore
new file mode 100644
index 0000000..dcbd87a
--- /dev/null
+++ b/.dockerignore
@@ -0,0 +1,13 @@
+.git
+.gitignore
+.DS_Store
+__pycache__
+*.pyc
+.pytest_cache
+.venv*
+
+android
+sentence_api/data/v
+sentence_api/data/work
+sentence_api/data/attempts
+sentence_api/data/oral_trainer.sqlite3*
diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000..98f37cb
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,9 @@
+.DS_Store
+__pycache__/
+*.py[cod]
+.pytest_cache/
+
+sentence_api/data/oral_trainer.sqlite3*
+sentence_api/data/v/
+sentence_api/data/work/
+sentence_api/data/attempts/
diff --git a/README.md b/README.md
index 5337648..90a7e4b 100644
--- a/README.md
+++ b/README.md
@@ -122,10 +122,11 @@ SDK 基于 AndroidX Media3 / ExoPlayer，支持在线视频流、本地缓存、
 
 详见 [android/README.md](android/README.md)。
 
-## 句子边界 API
+## 口语训练视频服务
 
-仓库新增了 `sentence_api/` 服务。它复用桌面播放器的静音检测算法，提前为视频
-生成句子边界，并根据视频 SHA-256 哈希向 Android 返回毫秒级的
-`start_ms` / `end_ms` 列表。服务不会在移动端请求时临时分析视频。
+`sentence_api/` 已扩展为完整的视频服务：支持超过 1 GB 的流式上传、MOSS 自动
+转写、句子文本校对、HTTP Range 在线播放、Android 视频目录、学生录音上传，以及
+内容正确度、有效语音时长、停顿和语速评分。原有的 SHA-256 句子边界接口保持兼容。
 
-详见 [sentence_api/README.md](sentence_api/README.md)。
+本地使用见 [sentence_api/README.md](sentence_api/README.md)，生产部署见
+[sentence_api/DEPLOYMENT.md](sentence_api/DEPLOYMENT.md)。
diff --git a/android/README.md b/android/README.md
index 27a6473..44ca8f5 100644
--- a/android/README.md
+++ b/android/README.md
@@ -2,7 +2,7 @@
 
 Android SDK for online oral-training video playback. It provides an embeddable
 Media3/ExoPlayer player, streaming cache, tablet gestures, sentence navigation,
-and a placeholder interface for future imitation-quality scoring.
+the server video catalog, and remote imitation-quality assessment.
 
 ## Modules
 
@@ -93,11 +93,36 @@ Override the service only when a staging or private deployment is required:
 OralTrainerSdk.init(
     context,
     OralTrainerSdkConfig(
-        sentenceBoundaryApiBaseUrl = "https://video_service.d1kt.cn"
+        sentenceBoundaryApiBaseUrl = "https://video_service.d1kt.cn",
+        assessmentApiKey = BuildConfig.ORAL_TRAINER_CLIENT_KEY,
     )
 )
 ```
 
+## Video Catalog
+
+The SDK loads ready videos from `GET /api/v1/videos` and resolves both absolute
+and relative stream URLs:
+
+```kotlin
+sdk.videoCatalogApi.fetch(object : VideoCatalogCallback {
+    override fun onSuccess(videos: List<TrainingVideoSummary>) {
+        val video = videos.first()
+        sdk.sentenceBoundaryApi.fetch(video.videoHash, object : SentenceBoundaryApiCallback {
+            override fun onSuccess(result: SentenceBoundaryApiResult) {
+                controller.loadItem(video.toTrainingMediaItem(result.sentences))
+            }
+
+            override fun onError(error: Throwable) {
+                controller.loadItem(video.toTrainingMediaItem())
+            }
+        })
+    }
+
+    override fun onError(error: Throwable) = Unit
+})
+```
+
 ## Local Video Testing
 
 The computer path `/Users/...` is not visible to an Android device. For a
@@ -110,19 +135,40 @@ service and pass its URL as `TrainingMediaItem.uri`. The SDK streams it and
 caches downloaded ranges locally. Copying large 4K files to each device is
 better reserved for explicitly offline courses.
 
-## Future Imitation Scoring
+## Remote Imitation Scoring
 
-Provide an implementation of `ImitationQualityAssessor` when the speech
-assessment algorithm is ready:
+Create the controller with the built-in remote assessor:
 
 ```kotlin
+val remoteAssessor = sdk.createRemoteImitationQualityAssessor()
 val controller = sdk.createController(
-    imitationAssessor = MyImitationQualityAssessor()
+    imitationAssessor = remoteAssessor
 )
 ```
 
 Then call `assessCurrentSentence(recordingUri, callback)` after the student
-records a sentence.
+records a sentence. The current `TrainingMediaItem.id` must be the server video
+SHA-256, which is already true for items created by `TrainingVideoSummary`:
+
+```kotlin
+controller.assessCurrentSentence(
+    recordingUri = recordingUri,
+    locale = "en",
+    callback = object : ImitationAssessmentCallback {
+        override fun onResult(result: ImitationAssessmentResult) {
+            val passed = result.passed == true
+            val score = result.overallScore
+            val durationRatio = result.durationRatio
+        }
+
+        override fun onError(error: Throwable) = Unit
+    }
+)
+```
+
+The server returns content, completeness, fluency, duration, pause, and speech
+rate scores. Phoneme pronunciation and prosody are nullable until the dedicated
+models are enabled.
 
 ## Build
 
diff --git a/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/ImitationQualityAssessor.kt b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/ImitationQualityAssessor.kt
index c3df927..2b4aefb 100644
--- a/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/ImitationQualityAssessor.kt
+++ b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/ImitationQualityAssessor.kt
@@ -18,6 +18,17 @@ data class ImitationAssessmentResult @JvmOverloads constructor(
     val completenessScore: Float? = null,
     val feedback: String? = null,
     val details: Map<String, String> = emptyMap(),
+    val passed: Boolean? = null,
+    val contentScore: Float? = null,
+    val durationScore: Float? = null,
+    val pauseScore: Float? = null,
+    val speechRateScore: Float? = null,
+    val durationRatio: Float? = null,
+    val referenceText: String? = null,
+    val recognizedText: String? = null,
+    val missingTokens: List<String> = emptyList(),
+    val extraTokens: List<String> = emptyList(),
+    val substitutions: List<Pair<String, String>> = emptyList(),
 )
 
 fun interface CancellableAssessment {
diff --git a/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/OralTrainerPlayerView.kt b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/OralTrainerPlayerView.kt
index 43c4a00..ae4b785 100644
--- a/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/OralTrainerPlayerView.kt
+++ b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/OralTrainerPlayerView.kt
@@ -4,6 +4,7 @@ import android.content.Context
 import android.util.AttributeSet
 import android.view.GestureDetector
 import android.view.MotionEvent
+import android.view.ViewConfiguration
 import android.widget.FrameLayout
 import androidx.media3.ui.PlayerView
 import kotlin.math.abs
@@ -21,6 +22,23 @@ class OralTrainerPlayerView @JvmOverloads constructor(
     private var controller: OralTrainerController? = null
     private var gestureControls = GestureControlsConfig()
     private val density = resources.displayMetrics.density
+    private val touchSlop = ViewConfiguration.get(context).scaledTouchSlop
+    private val longPressTimeoutMs = ViewConfiguration.getLongPressTimeout().toLong()
+    private var longPressStartX = 0f
+    private var longPressStartY = 0f
+    private var longPressActive = false
+    private var speedBeforeLongPress: Float? = null
+
+    private val longPressSpeedRunnable = Runnable {
+        val activeController = controller ?: return@Runnable
+        if (!gestureControls.enabled || !gestureControls.longPressAdjustsSpeed || width <= 0) {
+            return@Runnable
+        }
+        longPressActive = true
+        speedBeforeLongPress = activeController.snapshot().playbackSpeed
+        activeController.setPlaybackSpeed(speedForTouchX(longPressStartX))
+        activeController.dispatchGesture(GestureEvent(GestureKind.LONG_PRESS_SPEED))
+    }
 
     private val gestureDetector = GestureDetector(
         context,
@@ -99,6 +117,72 @@ class OralTrainerPlayerView @JvmOverloads constructor(
     }
 
     override fun dispatchTouchEvent(ev: MotionEvent): Boolean {
+        if (handleLongPressSpeed(ev)) {
+            return true
+        }
         return gestureDetector.onTouchEvent(ev) || super.dispatchTouchEvent(ev)
     }
+
+    override fun onDetachedFromWindow() {
+        cancelLongPressSpeed(restoreSpeed = true)
+        super.onDetachedFromWindow()
+    }
+
+    private fun handleLongPressSpeed(event: MotionEvent): Boolean {
+        if (!gestureControls.enabled || !gestureControls.longPressAdjustsSpeed) {
+            return false
+        }
+        when (event.actionMasked) {
+            MotionEvent.ACTION_DOWN -> {
+                longPressStartX = event.x
+                longPressStartY = event.y
+                longPressActive = false
+                speedBeforeLongPress = null
+                removeCallbacks(longPressSpeedRunnable)
+                postDelayed(longPressSpeedRunnable, longPressTimeoutMs)
+                return false
+            }
+
+            MotionEvent.ACTION_MOVE -> {
+                if (longPressActive) {
+                    controller?.setPlaybackSpeed(speedForTouchX(event.x))
+                    return true
+                }
+                val dx = abs(event.x - longPressStartX)
+                val dy = abs(event.y - longPressStartY)
+                if (dx > touchSlop || dy > touchSlop) {
+                    removeCallbacks(longPressSpeedRunnable)
+                }
+                return false
+            }
+
+            MotionEvent.ACTION_UP,
+            MotionEvent.ACTION_CANCEL -> {
+                val wasActive = longPressActive
+                cancelLongPressSpeed(restoreSpeed = wasActive)
+                return wasActive
+            }
+        }
+        return false
+    }
+
+    private fun cancelLongPressSpeed(restoreSpeed: Boolean) {
+        removeCallbacks(longPressSpeedRunnable)
+        if (restoreSpeed) {
+            controller?.setPlaybackSpeed(speedBeforeLongPress ?: 1f)
+        }
+        longPressActive = false
+        speedBeforeLongPress = null
+    }
+
+    private fun speedForTouchX(x: Float): Float {
+        val center = width / 2f
+        val distanceRatio = if (center <= 0f) 0f else abs(x - center) / center
+        val farFromCenter = distanceRatio >= 0.5f
+        return if (x < center) {
+            if (farFromCenter) gestureControls.longPressSlowerSpeed else gestureControls.longPressSlowSpeed
+        } else {
+            if (farFromCenter) gestureControls.longPressFasterSpeed else gestureControls.longPressFastSpeed
+        }
+    }
 }
diff --git a/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/OralTrainerSdk.kt b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/OralTrainerSdk.kt
index a1f78b0..26e4a95 100644
--- a/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/OralTrainerSdk.kt
+++ b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/OralTrainerSdk.kt
@@ -1,6 +1,7 @@
 package cn.learningpad.oraltrainer.sdk
 
 import android.content.Context
+import java.util.concurrent.CopyOnWriteArrayList
 
 class OralTrainerSdk private constructor(
     context: Context,
@@ -10,6 +11,12 @@ class OralTrainerSdk private constructor(
 
     val cache: OralTrainerCache = OralTrainerCache(appContext, config)
     val sentenceBoundaryApi: SentenceBoundaryApi = SentenceBoundaryApi(appContext, config)
+    val videoCatalogApi: VideoCatalogApi = VideoCatalogApi(appContext, config)
+    private val remoteAssessors = CopyOnWriteArrayList<RemoteImitationQualityAssessor>()
+
+    fun createRemoteImitationQualityAssessor(): RemoteImitationQualityAssessor {
+        return RemoteImitationQualityAssessor(appContext, config).also(remoteAssessors::add)
+    }
 
     @JvmOverloads
     fun createController(
@@ -26,6 +33,9 @@ class OralTrainerSdk private constructor(
 
     fun release() {
         sentenceBoundaryApi.release()
+        videoCatalogApi.release()
+        remoteAssessors.forEach { it.release() }
+        remoteAssessors.clear()
         StreamingCache.release()
     }
 
diff --git a/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/OralTrainerSdkConfig.kt b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/OralTrainerSdkConfig.kt
index b033443..92443f1 100644
--- a/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/OralTrainerSdkConfig.kt
+++ b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/OralTrainerSdkConfig.kt
@@ -9,4 +9,5 @@ data class OralTrainerSdkConfig @JvmOverloads constructor(
     val connectTimeoutMs: Int = 15_000,
     val readTimeoutMs: Int = 30_000,
     val sentenceBoundaryApiBaseUrl: String = "https://video_service.d1kt.cn",
+    val assessmentApiKey: String? = null,
 )
diff --git a/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/PlaybackModels.kt b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/PlaybackModels.kt
index f6f300d..fc1ea17 100644
--- a/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/PlaybackModels.kt
+++ b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/PlaybackModels.kt
@@ -27,6 +27,7 @@ enum class GestureKind {
     SINGLE_TAP,
     SWIPE_LEFT,
     SWIPE_RIGHT,
+    LONG_PRESS_SPEED,
 }
 
 interface OralTrainerListener {
diff --git a/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/PlayerConfig.kt b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/PlayerConfig.kt
index 6d66981..8ba91bc 100644
--- a/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/PlayerConfig.kt
+++ b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/PlayerConfig.kt
@@ -16,6 +16,11 @@ data class GestureControlsConfig @JvmOverloads constructor(
     val rightSwipeAction: SwipeAction = SwipeAction.NEXT_SENTENCE_OR_FORWARD,
     val minSwipeDistanceDp: Float = 48f,
     val minSwipeVelocityDpPerSecond: Float = 160f,
+    val longPressAdjustsSpeed: Boolean = true,
+    val longPressSlowSpeed: Float = 0.75f,
+    val longPressSlowerSpeed: Float = 0.5f,
+    val longPressFastSpeed: Float = 1.5f,
+    val longPressFasterSpeed: Float = 2.0f,
 )
 
 enum class SwipeAction {
diff --git a/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/RemoteImitationQualityAssessor.kt b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/RemoteImitationQualityAssessor.kt
new file mode 100644
index 0000000..652e8c6
--- /dev/null
+++ b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/RemoteImitationQualityAssessor.kt
@@ -0,0 +1,198 @@
+package cn.learningpad.oraltrainer.sdk
+
+import android.content.ContentResolver
+import android.content.Context
+import android.net.Uri
+import android.os.Handler
+import android.os.Looper
+import org.json.JSONArray
+import org.json.JSONObject
+import java.io.BufferedInputStream
+import java.io.BufferedOutputStream
+import java.io.IOException
+import java.net.HttpURLConnection
+import java.net.URL
+import java.util.UUID
+import java.util.concurrent.ExecutorService
+import java.util.concurrent.Executors
+import java.util.concurrent.Future
+
+/** Uploads one sentence recording to the server-side MOSS assessment endpoint. */
+class RemoteImitationQualityAssessor internal constructor(
+    context: Context,
+    private val config: OralTrainerSdkConfig,
+) : ImitationQualityAssessor {
+    private val appContext = context.applicationContext
+    private val contentResolver: ContentResolver = appContext.contentResolver
+    private val executor: ExecutorService = Executors.newCachedThreadPool()
+    private val mainHandler = Handler(Looper.getMainLooper())
+    private val baseUri = Uri.parse(config.sentenceBoundaryApiBaseUrl.trimEnd('/'))
+
+    init {
+        require(baseUri.scheme == "https" || baseUri.scheme == "http") {
+            "sentenceBoundaryApiBaseUrl must use http or https."
+        }
+        require(!baseUri.host.isNullOrBlank()) {
+            "sentenceBoundaryApiBaseUrl must include a host."
+        }
+    }
+
+    override fun assess(
+        request: ImitationAssessmentRequest,
+        callback: ImitationAssessmentCallback,
+    ): CancellableAssessment {
+        val future = executor.submit {
+            try {
+                val result = assessBlocking(request)
+                if (!Thread.currentThread().isInterrupted) {
+                    mainHandler.post { callback.onResult(result) }
+                }
+            } catch (error: Throwable) {
+                if (!Thread.currentThread().isInterrupted) {
+                    mainHandler.post { callback.onError(error) }
+                }
+            }
+        }
+        return CancellableAssessment { future.cancel(true) }
+    }
+
+    fun release() {
+        executor.shutdownNow()
+    }
+
+    private fun assessBlocking(request: ImitationAssessmentRequest): ImitationAssessmentResult {
+        val videoHash = request.metadata["video_hash"] ?: request.mediaId
+        require(SHA256_PATTERN.matches(videoHash)) {
+            "ImitationAssessmentRequest.mediaId or metadata[video_hash] must be a SHA-256 hash."
+        }
+        val endpoint = baseUri.buildUpon()
+            .appendPath("api")
+            .appendPath("v1")
+            .appendPath("videos")
+            .appendPath(videoHash.lowercase())
+            .appendPath("sentences")
+            .appendPath(request.sentence.index.toString())
+            .appendPath("assessments")
+            .build()
+        val connection = URL(endpoint.toString()).openConnection() as HttpURLConnection
+        val boundary = "----OralTrainer-${UUID.randomUUID()}"
+        return try {
+            connection.requestMethod = "POST"
+            connection.doOutput = true
+            connection.connectTimeout = config.connectTimeoutMs
+            connection.readTimeout = maxOf(config.readTimeoutMs, 180_000)
+            connection.setChunkedStreamingMode(64 * 1024)
+            connection.setRequestProperty("Accept", "application/json")
+            connection.setRequestProperty("User-Agent", config.userAgent)
+            connection.setRequestProperty("Content-Type", "multipart/form-data; boundary=$boundary")
+            config.assessmentApiKey?.takeIf { it.isNotBlank() }?.let {
+                connection.setRequestProperty("X-Client-Key", it)
+            }
+            BufferedOutputStream(connection.outputStream).use { output ->
+                writeFormField(output, boundary, "language", request.locale ?: request.sentence.language.orEmpty())
+                writeFilePart(output, boundary, request.recordingUri)
+                output.write("--$boundary--\r\n".toByteArray(Charsets.UTF_8))
+            }
+            val statusCode = connection.responseCode
+            val stream = if (statusCode in 200..299) connection.inputStream else connection.errorStream
+            val body = stream?.bufferedReader(Charsets.UTF_8)?.use { it.readText() }.orEmpty()
+            if (statusCode !in 200..299) {
+                throw RemoteAssessmentException(statusCode, body)
+            }
+            parseResponse(body)
+        } finally {
+            connection.disconnect()
+        }
+    }
+
+    private fun writeFormField(output: BufferedOutputStream, boundary: String, name: String, value: String) {
+        output.write("--$boundary\r\n".toByteArray(Charsets.UTF_8))
+        output.write("Content-Disposition: form-data; name=\"$name\"\r\n\r\n".toByteArray(Charsets.UTF_8))
+        output.write(value.toByteArray(Charsets.UTF_8))
+        output.write("\r\n".toByteArray(Charsets.UTF_8))
+    }
+
+    private fun writeFilePart(output: BufferedOutputStream, boundary: String, uri: Uri) {
+        val fileName = uri.lastPathSegment?.substringAfterLast('/')
+            ?.replace("\r", "_")
+            ?.replace("\n", "_")
+            ?.replace("\"", "_")
+            ?.ifBlank { "recording.wav" }
+            ?: "recording.wav"
+        val mimeType = contentResolver.getType(uri) ?: "audio/wav"
+        output.write("--$boundary\r\n".toByteArray(Charsets.UTF_8))
+        output.write(
+            "Content-Disposition: form-data; name=\"audio\"; filename=\"$fileName\"\r\n".toByteArray(
+                Charsets.UTF_8
+            )
+        )
+        output.write("Content-Type: $mimeType\r\n\r\n".toByteArray(Charsets.UTF_8))
+        contentResolver.openInputStream(uri)?.use { input ->
+            BufferedInputStream(input).use { buffered ->
+                val buffer = ByteArray(64 * 1024)
+                while (true) {
+                    if (Thread.currentThread().isInterrupted) throw IOException("Assessment request was cancelled.")
+                    val count = buffered.read(buffer)
+                    if (count < 0) break
+                    output.write(buffer, 0, count)
+                }
+            }
+        } ?: throw IOException("Cannot open recording URI: $uri")
+        output.write("\r\n".toByteArray(Charsets.UTF_8))
+    }
+
+    private fun parseResponse(body: String): ImitationAssessmentResult {
+        val root = JSONObject(body)
+        val substitutions = buildList {
+            val array = root.optJSONArray("substitutions") ?: JSONArray()
+            for (index in 0 until array.length()) {
+                val item = array.optJSONObject(index) ?: continue
+                add(item.optString("expected") to item.optString("actual"))
+            }
+        }
+        val details = buildMap {
+            val objectValue = root.optJSONObject("details") ?: return@buildMap
+            for (key in objectValue.keys()) put(key, objectValue.optString(key))
+        }
+        return ImitationAssessmentResult(
+            overallScore = root.getDouble("overall_score").toFloat(),
+            pronunciationScore = root.optNullableDouble("pronunciation_score")?.toFloat(),
+            fluencyScore = root.optNullableDouble("fluency_score")?.toFloat(),
+            completenessScore = root.optNullableDouble("completeness_score")?.toFloat(),
+            feedback = root.optString("feedback").ifBlank { null },
+            details = details,
+            passed = if (root.has("passed")) root.optBoolean("passed") else null,
+            contentScore = root.optNullableDouble("content_score")?.toFloat(),
+            durationScore = root.optNullableDouble("duration_score")?.toFloat(),
+            pauseScore = root.optNullableDouble("pause_score")?.toFloat(),
+            speechRateScore = root.optNullableDouble("speech_rate_score")?.toFloat(),
+            durationRatio = root.optNullableDouble("duration_ratio")?.toFloat(),
+            referenceText = root.optString("reference_text").ifBlank { null },
+            recognizedText = root.optString("recognized_text").ifBlank { null },
+            missingTokens = root.optStringArray("missing_tokens"),
+            extraTokens = root.optStringArray("extra_tokens"),
+            substitutions = substitutions,
+        )
+    }
+
+    companion object {
+        private val SHA256_PATTERN = Regex("^[0-9a-fA-F]{64}$")
+    }
+}
+
+class RemoteAssessmentException(
+    val statusCode: Int,
+    responseBody: String,
+) : IOException("Remote assessment returned HTTP $statusCode: $responseBody")
+
+private fun JSONObject.optNullableDouble(name: String): Double? {
+    if (!has(name) || isNull(name)) return null
+    return optDouble(name).takeUnless { it.isNaN() }
+}
+
+private fun JSONObject.optStringArray(name: String): List<String> {
+    val array = optJSONArray(name) ?: return emptyList()
+    return buildList(array.length()) {
+        for (index in 0 until array.length()) add(array.optString(index))
+    }
+}
diff --git a/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/SentenceBoundaryApi.kt b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/SentenceBoundaryApi.kt
index 56a232f..0da60e6 100644
--- a/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/SentenceBoundaryApi.kt
+++ b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/SentenceBoundaryApi.kt
@@ -147,6 +147,12 @@ class SentenceBoundaryApi internal constructor(
                         startMs = sentence.getLong("start_ms"),
                         endMs = sentence.getLong("end_ms"),
                         text = if (sentence.isNull("text")) null else sentence.getString("text"),
+                        language = if (sentence.isNull("language")) null else sentence.getString("language"),
+                        referenceSpeechDurationMs = if (sentence.isNull("reference_speech_duration_ms")) {
+                            null
+                        } else {
+                            sentence.getLong("reference_speech_duration_ms")
+                        },
                     )
                 )
             }
diff --git a/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/TrainingModels.kt b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/TrainingModels.kt
index 2f5bded..736eb4e 100644
--- a/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/TrainingModels.kt
+++ b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/TrainingModels.kt
@@ -8,11 +8,16 @@ data class SentenceBoundary @JvmOverloads constructor(
     val startMs: Long,
     val endMs: Long,
     val text: String? = null,
+    val language: String? = null,
+    val referenceSpeechDurationMs: Long? = null,
 ) {
     init {
         require(index >= 0) { "Sentence index must be non-negative." }
         require(startMs >= 0) { "Sentence startMs must be non-negative." }
         require(endMs > startMs) { "Sentence endMs must be greater than startMs." }
+        require(referenceSpeechDurationMs == null || referenceSpeechDurationMs > 0) {
+            "Sentence referenceSpeechDurationMs must be positive when present."
+        }
     }
 }
 
diff --git a/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/VideoCatalogApi.kt b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/VideoCatalogApi.kt
new file mode 100644
index 0000000..9b58eb3
--- /dev/null
+++ b/android/oral-trainer-sdk/src/main/java/cn/learningpad/oraltrainer/sdk/VideoCatalogApi.kt
@@ -0,0 +1,116 @@
+package cn.learningpad.oraltrainer.sdk
+
+import android.content.Context
+import android.net.Uri
+import android.os.Handler
+import android.os.Looper
+import org.json.JSONObject
+import java.io.IOException
+import java.net.HttpURLConnection
+import java.net.URL
+import java.util.concurrent.ExecutorService
+import java.util.concurrent.Executors
+
+data class TrainingVideoSummary(
+    val videoHash: String,
+    val title: String,
+    val streamUri: Uri,
+    val durationMs: Long?,
+    val sizeBytes: Long,
+    val language: String?,
+    val sentenceCount: Int,
+    val status: String,
+) {
+    fun toTrainingMediaItem(sentences: List<SentenceBoundary> = emptyList()): TrainingMediaItem {
+        return TrainingMediaItem(
+            id = videoHash,
+            title = title,
+            uri = streamUri,
+            sentences = sentences,
+            customCacheKey = videoHash,
+        )
+    }
+}
+
+interface VideoCatalogCallback {
+    fun onSuccess(videos: List<TrainingVideoSummary>)
+
+    fun onError(error: Throwable)
+}
+
+class VideoCatalogApi internal constructor(
+    context: Context,
+    private val config: OralTrainerSdkConfig,
+) {
+    private val mainHandler = Handler(Looper.getMainLooper())
+    private val executor: ExecutorService = Executors.newCachedThreadPool()
+    private val baseUrl = config.sentenceBoundaryApiBaseUrl.trimEnd('/') + "/"
+
+    fun fetch(callback: VideoCatalogCallback): CancellableRequest {
+        val future = executor.submit {
+            try {
+                val videos = fetchBlocking()
+                if (!Thread.currentThread().isInterrupted) {
+                    mainHandler.post { callback.onSuccess(videos) }
+                }
+            } catch (error: Throwable) {
+                if (!Thread.currentThread().isInterrupted) {
+                    mainHandler.post { callback.onError(error) }
+                }
+            }
+        }
+        return CancellableRequest(future)
+    }
+
+    fun release() {
+        executor.shutdownNow()
+    }
+
+    private fun fetchBlocking(): List<TrainingVideoSummary> {
+        val endpoint = URL(URL(baseUrl), "api/v1/videos")
+        val connection = endpoint.openConnection() as HttpURLConnection
+        return try {
+            connection.requestMethod = "GET"
+            connection.connectTimeout = config.connectTimeoutMs
+            connection.readTimeout = config.readTimeoutMs
+            connection.setRequestProperty("Accept", "application/json")
+            connection.setRequestProperty("User-Agent", config.userAgent)
+            val statusCode = connection.responseCode
+            val stream = if (statusCode in 200..299) connection.inputStream else connection.errorStream
+            val body = stream?.bufferedReader(Charsets.UTF_8)?.use { it.readText() }.orEmpty()
+            if (statusCode !in 200..299) throw VideoCatalogApiException(statusCode, body)
+            parseResponse(body)
+        } finally {
+            connection.disconnect()
+        }
+    }
+
+    private fun parseResponse(body: String): List<TrainingVideoSummary> {
+        val videos = JSONObject(body).getJSONArray("videos")
+        return buildList(videos.length()) {
+            for (index in 0 until videos.length()) {
+                val item = videos.getJSONObject(index)
+                val status = item.getString("status")
+                if (status != "ready") continue
+                val streamUrl = URL(URL(baseUrl), item.getString("stream_url"))
+                add(
+                    TrainingVideoSummary(
+                        videoHash = item.getString("video_hash"),
+                        title = item.getString("title"),
+                        streamUri = Uri.parse(streamUrl.toString()),
+                        durationMs = if (item.isNull("duration_ms")) null else item.getLong("duration_ms"),
+                        sizeBytes = item.getLong("size_bytes"),
+                        language = if (item.isNull("language")) null else item.getString("language"),
+                        sentenceCount = item.getInt("sentence_count"),
+                        status = status,
+                    )
+                )
+            }
+        }
+    }
+}
+
+class VideoCatalogApiException(
+    val statusCode: Int,
+    responseBody: String,
+) : IOException("Video catalog API returned HTTP $statusCode: $responseBody")
diff --git a/android/sample-app/src/main/AndroidManifest.xml b/android/sample-app/src/main/AndroidManifest.xml
index ab68f66..ee49011 100644
--- a/android/sample-app/src/main/AndroidManifest.xml
+++ b/android/sample-app/src/main/AndroidManifest.xml
@@ -3,7 +3,7 @@
 
     <application
         android:allowBackup="true"
-        android:label="Oral Trainer Sample"
+        android:label="@string/app_name"
         android:supportsRtl="true"
         android:theme="@style/AppTheme">
         <activity
diff --git a/android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/MainActivity.kt b/android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/MainActivity.kt
index 40a00a8..5a5e706 100644
--- a/android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/MainActivity.kt
+++ b/android/sample-app/src/main/java/cn/learningpad/oraltrainer/sample/MainActivity.kt
@@ -2,16 +2,24 @@ package cn.learningpad.oraltrainer.sample
 
 import android.app.Activity
 import android.content.Intent
+import android.content.res.ColorStateList
 import android.graphics.Color
+import android.graphics.Typeface
+import android.graphics.drawable.GradientDrawable
 import android.net.Uri
 import android.os.Bundle
 import android.view.Gravity
+import android.view.View
 import android.view.ViewGroup
 import android.widget.Button
+import android.widget.FrameLayout
+import android.widget.HorizontalScrollView
 import android.widget.LinearLayout
+import android.widget.ProgressBar
 import android.widget.TextView
 import androidx.media3.common.MimeTypes
 import cn.learningpad.oraltrainer.sdk.GestureEvent
+import cn.learningpad.oraltrainer.sdk.GestureKind
 import cn.learningpad.oraltrainer.sdk.LoopMode
 import cn.learningpad.oraltrainer.sdk.OralTrainerController
 import cn.learningpad.oraltrainer.sdk.OralTrainerListener
@@ -21,21 +29,33 @@ import cn.learningpad.oraltrainer.sdk.PlaybackSnapshot
 import cn.learningpad.oraltrainer.sdk.PlayerConfig
 import cn.learningpad.oraltrainer.sdk.SentenceBoundary
 import cn.learningpad.oraltrainer.sdk.SentenceBoundaryApiCallback
-import cn.learningpad.oraltrainer.sdk.TrainingMediaItem
 import cn.learningpad.oraltrainer.sdk.SentenceBoundaryApiResult
+import cn.learningpad.oraltrainer.sdk.TrainingMediaItem
+import cn.learningpad.oraltrainer.sdk.TrainingVideoSummary
+import cn.learningpad.oraltrainer.sdk.VideoCatalogCallback
+import java.util.Locale
 
 class MainActivity : Activity() {
-    private companion object {
-        const val PICK_VIDEO_REQUEST = 1001
-    }
-
     private lateinit var sdk: OralTrainerSdk
     private lateinit var controller: OralTrainerController
+    private lateinit var playerView: OralTrainerPlayerView
+    private lateinit var lessonTitleText: TextView
     private lateinit var statusText: TextView
     private lateinit var sentenceText: TextView
+    private lateinit var sentenceMetaText: TextView
+    private lateinit var timeText: TextView
+    private lateinit var speedText: TextView
+    private lateinit var progressBar: ProgressBar
+    private lateinit var catalogList: LinearLayout
+    private lateinit var catalogStatusText: TextView
+
+    private var activeItemId: String = SAMPLE_ID
+    private var currentSentenceCount = 0
+    private var catalogVideos: List<TrainingVideoSummary> = emptyList()
 
     override fun onCreate(savedInstanceState: Bundle?) {
         super.onCreate(savedInstanceState)
+        configureWindow()
 
         sdk = OralTrainerSdk.init(this)
         controller = sdk.createController(
@@ -47,77 +67,10 @@ class MainActivity : Activity() {
         )
         controller.setLoopMode(LoopMode.ALL)
 
-        val playerView = OralTrainerPlayerView(this).apply {
-            bind(controller)
-            layoutParams = LinearLayout.LayoutParams(
-                ViewGroup.LayoutParams.MATCH_PARENT,
-                0,
-                1f,
-            )
-        }
-
-        statusText = TextView(this).apply {
-            setTextColor(Color.WHITE)
-            textSize = 14f
-            setPadding(24, 16, 24, 8)
-        }
-        sentenceText = TextView(this).apply {
-            setTextColor(Color.WHITE)
-            textSize = 18f
-            setPadding(24, 4, 24, 16)
-        }
-
-        val controls = LinearLayout(this).apply {
-            orientation = LinearLayout.HORIZONTAL
-            gravity = Gravity.CENTER
-            setPadding(16, 8, 16, 20)
-            addView(commandButton("快退") { controller.rewind() })
-            addView(commandButton("播放/暂停") { controller.togglePlayPause() })
-            addView(commandButton("快进") { controller.fastForward() })
-            addView(commandButton("上一句") { controller.previousSentenceOrRewind() })
-            addView(commandButton("下一句") { controller.nextSentenceOrForward() })
-        }
-
-        val root = LinearLayout(this).apply {
-            orientation = LinearLayout.VERTICAL
-            setBackgroundColor(Color.rgb(18, 18, 18))
-            addView(playerView)
-            addView(statusText)
-            addView(sentenceText)
-            addView(controls)
-            addView(Button(this@MainActivity).apply {
-                text = "选择本地视频"
-                setOnClickListener { openLocalVideoPicker() }
-            })
-        }
-        setContentView(root)
-
-        controller.addListener(object : OralTrainerListener {
-            override fun onPlaybackSnapshot(snapshot: PlaybackSnapshot) {
-                statusText.text = buildString {
-                    append(if (snapshot.isPlaying) "播放中" else "已暂停")
-                    append("  ")
-                    append(format(snapshot.positionMs))
-                    append(" / ")
-                    append(format(snapshot.durationMs))
-                    append("  速度 ")
-                    append(snapshot.playbackSpeed)
-                    append("x")
-                }
-            }
-
-            override fun onSentenceChanged(sentence: SentenceBoundary?) {
-                sentenceText.text = sentence?.let {
-                    "第 ${it.index + 1} 句：${it.text.orEmpty()}"
-                } ?: "当前没有句子边界"
-            }
-
-            override fun onGesture(event: GestureEvent) {
-                statusText.text = "手势：${event.kind} ${event.action ?: ""}"
-            }
-        })
-
-        controller.loadItem(sampleOnlineLesson())
+        setContentView(createContentView())
+        bindPlayerEvents()
+        loadSampleLesson()
+        loadCatalog()
     }
 
     override fun onActivityResult(requestCode: Int, resultCode: Int, data: Intent?) {
@@ -132,21 +85,38 @@ class MainActivity : Activity() {
                 contentResolver.takePersistableUriPermission(uri, persistableFlags)
             }
         }
+
         val item = TrainingMediaItem(
             id = uri.toString(),
             title = uri.lastPathSegment ?: "本地视频",
             uri = uri,
         )
+        activeItemId = item.id
+        currentSentenceCount = 0
         controller.loadItem(item)
-        statusText.text = "正在获取句子边界..."
+        lessonTitleText.text = item.title
+        sentenceMetaText.text = "句子边界分析中"
+        sentenceText.text = "正在匹配服务端句子边界"
+        statusText.text = "正在获取本地视频的句子边界..."
+        renderCatalog()
+
         sdk.sentenceBoundaryApi.fetchForUri(uri, contentResolver, object : SentenceBoundaryApiCallback {
             override fun onSuccess(result: SentenceBoundaryApiResult) {
+                if (activeItemId != item.id) {
+                    return
+                }
+                currentSentenceCount = result.sentences.size
                 controller.loadItem(item.copy(sentences = result.sentences))
                 statusText.text = "句子边界已加载：${result.sentences.size} 句"
             }
 
             override fun onError(error: Throwable) {
-                statusText.text = "句子边界获取失败，使用 10 秒跳转：${error.message.orEmpty()}"
+                if (activeItemId != item.id) {
+                    return
+                }
+                statusText.text = "句子边界获取失败，使用默认快退/快进：${error.message.orEmpty()}"
+                sentenceMetaText.text = "暂无句子边界"
+                sentenceText.text = item.title
             }
         })
     }
@@ -156,18 +126,381 @@ class MainActivity : Activity() {
         super.onDestroy()
     }
 
-    private fun commandButton(label: String, action: () -> Unit): Button {
+    @Suppress("DEPRECATION")
+    private fun configureWindow() {
+        window.statusBarColor = COLOR_BACKGROUND
+        window.navigationBarColor = COLOR_BACKGROUND
+    }
+
+    private fun createContentView(): View {
+        return LinearLayout(this).apply {
+            orientation = LinearLayout.VERTICAL
+            setBackgroundColor(COLOR_BACKGROUND)
+            addView(createHeader())
+            addView(createPlayerSection())
+            addView(createSentenceSection())
+            addView(createCatalogSection())
+        }
+    }
+
+    private fun createHeader(): View {
+        val titleBlock = LinearLayout(this).apply {
+            orientation = LinearLayout.VERTICAL
+            gravity = Gravity.CENTER_VERTICAL
+            layoutParams = LinearLayout.LayoutParams(0, ViewGroup.LayoutParams.WRAP_CONTENT, 1f)
+            addView(TextView(this@MainActivity).apply {
+                text = "口语宝"
+                setTextColor(Color.WHITE)
+                textSize = 26f
+                typeface = Typeface.DEFAULT_BOLD
+                includeFontPadding = false
+            })
+            addView(TextView(this@MainActivity).apply {
+                text = "中英跟读训练"
+                setTextColor(COLOR_TEXT_MUTED)
+                textSize = 13f
+                setPadding(1.dp, 4.dp, 0, 0)
+            })
+        }
+
+        return LinearLayout(this).apply {
+            orientation = LinearLayout.HORIZONTAL
+            gravity = Gravity.CENTER_VERTICAL
+            setPadding(18.dp, 16.dp, 18.dp, 12.dp)
+            addView(titleBlock)
+            addView(headerButton("导入") { openLocalVideoPicker() })
+            addView(headerButton("刷新") { loadCatalog() })
+        }
+    }
+
+    private fun createPlayerSection(): View {
+        playerView = OralTrainerPlayerView(this).apply {
+            bind(controller)
+            layoutParams = FrameLayout.LayoutParams(
+                ViewGroup.LayoutParams.MATCH_PARENT,
+                ViewGroup.LayoutParams.MATCH_PARENT,
+            )
+        }
+
+        val overlay = LinearLayout(this).apply {
+            orientation = LinearLayout.HORIZONTAL
+            gravity = Gravity.CENTER_VERTICAL
+            setPadding(12.dp, 12.dp, 12.dp, 12.dp)
+            layoutParams = FrameLayout.LayoutParams(
+                ViewGroup.LayoutParams.MATCH_PARENT,
+                ViewGroup.LayoutParams.WRAP_CONTENT,
+                Gravity.TOP,
+            )
+        }
+
+        timeText = overlayPill("--:-- / --:--").apply {
+            layoutParams = LinearLayout.LayoutParams(0, ViewGroup.LayoutParams.WRAP_CONTENT, 1f)
+        }
+        speedText = overlayPill("1x")
+        overlay.addView(timeText)
+        overlay.addView(speedText)
+
+        return FrameLayout(this).apply {
+            background = rounded(COLOR_SURFACE, 8f, COLOR_BORDER)
+            clipToOutline = true
+            layoutParams = LinearLayout.LayoutParams(
+                ViewGroup.LayoutParams.MATCH_PARENT,
+                0,
+                1f,
+            ).withMargins(16.dp, 0, 16.dp, 10.dp)
+            addView(playerView)
+            addView(overlay)
+        }
+    }
+
+    private fun createSentenceSection(): View {
+        lessonTitleText = TextView(this).apply {
+            setTextColor(COLOR_TEXT_DARK)
+            textSize = 17f
+            typeface = Typeface.DEFAULT_BOLD
+            maxLines = 1
+        }
+        sentenceMetaText = TextView(this).apply {
+            setTextColor(COLOR_ACCENT_DEEP)
+            textSize = 13f
+            setPadding(0, 8.dp, 0, 0)
+        }
+        sentenceText = TextView(this).apply {
+            setTextColor(COLOR_TEXT_DARK)
+            textSize = 22f
+            typeface = Typeface.DEFAULT_BOLD
+            setLineSpacing(2.dp.toFloat(), 1.05f)
+            setPadding(0, 10.dp, 0, 10.dp)
+        }
+        progressBar = ProgressBar(this, null, android.R.attr.progressBarStyleHorizontal).apply {
+            max = PROGRESS_MAX
+            progress = 0
+            progressTintList = ColorStateList.valueOf(COLOR_ACCENT)
+            progressBackgroundTintList = ColorStateList.valueOf(COLOR_PROGRESS_TRACK)
+            layoutParams = LinearLayout.LayoutParams(
+                ViewGroup.LayoutParams.MATCH_PARENT,
+                6.dp,
+            )
+        }
+        statusText = TextView(this).apply {
+            setTextColor(COLOR_TEXT_SUBTLE)
+            textSize = 13f
+            setPadding(0, 10.dp, 0, 0)
+        }
+
+        return LinearLayout(this).apply {
+            orientation = LinearLayout.VERTICAL
+            background = rounded(Color.WHITE, 8f, COLOR_LIGHT_BORDER)
+            setPadding(18.dp, 16.dp, 18.dp, 16.dp)
+            layoutParams = LinearLayout.LayoutParams(
+                ViewGroup.LayoutParams.MATCH_PARENT,
+                ViewGroup.LayoutParams.WRAP_CONTENT,
+            ).withMargins(16.dp, 0, 16.dp, 12.dp)
+            addView(lessonTitleText)
+            addView(sentenceMetaText)
+            addView(sentenceText)
+            addView(progressBar)
+            addView(statusText)
+        }
+    }
+
+    private fun createCatalogSection(): View {
+        catalogStatusText = TextView(this).apply {
+            setTextColor(COLOR_TEXT_MUTED)
+            textSize = 13f
+            gravity = Gravity.END
+            text = "正在同步"
+            layoutParams = LinearLayout.LayoutParams(0, ViewGroup.LayoutParams.WRAP_CONTENT, 1f)
+        }
+        catalogList = LinearLayout(this).apply {
+            orientation = LinearLayout.HORIZONTAL
+            setPadding(0, 10.dp, 0, 4.dp)
+        }
+
+        val header = LinearLayout(this).apply {
+            orientation = LinearLayout.HORIZONTAL
+            gravity = Gravity.CENTER_VERTICAL
+            addView(TextView(this@MainActivity).apply {
+                text = "课程"
+                setTextColor(Color.WHITE)
+                textSize = 16f
+                typeface = Typeface.DEFAULT_BOLD
+            })
+            addView(catalogStatusText)
+        }
+
+        val scroller = HorizontalScrollView(this).apply {
+            isHorizontalScrollBarEnabled = false
+            addView(catalogList)
+        }
+
+        return LinearLayout(this).apply {
+            orientation = LinearLayout.VERTICAL
+            setPadding(16.dp, 0, 16.dp, 14.dp)
+            addView(header)
+            addView(scroller)
+        }
+    }
+
+    private fun bindPlayerEvents() {
+        controller.addListener(object : OralTrainerListener {
+            override fun onPlaybackSnapshot(snapshot: PlaybackSnapshot) {
+                timeText.text = "${formatTime(snapshot.positionMs)} / ${formatTime(snapshot.durationMs)}"
+                speedText.text = formatSpeed(snapshot.playbackSpeed)
+                statusText.text = playbackStatus(snapshot)
+                progressBar.progress = playbackProgress(snapshot)
+            }
+
+            override fun onMediaChanged(item: TrainingMediaItem?) {
+                lessonTitleText.text = item?.title ?: "未选择课程"
+                currentSentenceCount = item?.sentences?.size ?: 0
+            }
+
+            override fun onSentenceChanged(sentence: SentenceBoundary?) {
+                if (sentence == null) {
+                    sentenceMetaText.text = "暂无句子边界"
+                    sentenceText.text = lessonTitleText.text
+                    return
+                }
+                val countText = if (currentSentenceCount > 0) {
+                    " / $currentSentenceCount"
+                } else {
+                    ""
+                }
+                sentenceMetaText.text = "第 ${sentence.index + 1}$countText 句  ${formatTime(sentence.startMs)}-${formatTime(sentence.endMs)}"
+                sentenceText.text = sentence.text?.takeIf { it.isNotBlank() } ?: "当前句子"
+            }
+
+            override fun onGesture(event: GestureEvent) {
+                statusText.text = when (event.kind) {
+                    GestureKind.SINGLE_TAP -> "播放状态已切换"
+                    GestureKind.SWIPE_LEFT -> "已跳到上一句"
+                    GestureKind.SWIPE_RIGHT -> "已跳到下一句"
+                    GestureKind.LONG_PRESS_SPEED -> "变速播放中"
+                }
+            }
+
+            override fun onPlayerError(error: Throwable) {
+                statusText.text = "播放失败：${error.message.orEmpty()}"
+            }
+        })
+    }
+
+    private fun loadSampleLesson() {
+        val item = sampleOnlineLesson()
+        activeItemId = item.id
+        currentSentenceCount = item.sentences.size
+        controller.loadItem(item)
+        lessonTitleText.text = item.title
+        renderCatalog()
+    }
+
+    private fun loadCatalog() {
+        catalogStatusText.text = "正在同步"
+        sdk.videoCatalogApi.fetch(object : VideoCatalogCallback {
+            override fun onSuccess(videos: List<TrainingVideoSummary>) {
+                catalogVideos = videos
+                catalogStatusText.text = if (videos.isEmpty()) {
+                    "暂无云端课程"
+                } else {
+                    "${videos.size} 个云端课程"
+                }
+                renderCatalog()
+            }
+
+            override fun onError(error: Throwable) {
+                catalogStatusText.text = "云端暂不可用"
+                renderCatalog()
+            }
+        })
+    }
+
+    private fun renderCatalog() {
+        if (!::catalogList.isInitialized) {
+            return
+        }
+        catalogList.removeAllViews()
+        catalogList.addView(
+            videoCard(
+                id = SAMPLE_ID,
+                title = "示例课程",
+                meta = "4 句 · ${formatTime(16_000L)}",
+            ) {
+                loadSampleLesson()
+            }
+        )
+        catalogVideos.forEach { video ->
+            catalogList.addView(
+                videoCard(
+                    id = video.videoHash,
+                    title = video.title,
+                    meta = "${video.sentenceCount} 句 · ${formatTime(video.durationMs ?: -1L)}",
+                ) {
+                    loadRemoteVideo(video)
+                }
+            )
+        }
+    }
+
+    private fun loadRemoteVideo(video: TrainingVideoSummary) {
+        activeItemId = video.videoHash
+        lessonTitleText.text = video.title
+        sentenceMetaText.text = "句子边界加载中"
+        sentenceText.text = video.title
+        statusText.text = "正在加载云端课程..."
+        renderCatalog()
+
+        val loadingItem = video.toTrainingMediaItem()
+        controller.loadItem(loadingItem)
+        currentSentenceCount = video.sentenceCount
+        sdk.sentenceBoundaryApi.fetch(video.videoHash, object : SentenceBoundaryApiCallback {
+            override fun onSuccess(result: SentenceBoundaryApiResult) {
+                if (activeItemId != video.videoHash) {
+                    return
+                }
+                currentSentenceCount = result.sentences.size
+                controller.loadItem(video.toTrainingMediaItem(result.sentences))
+                statusText.text = "课程已就绪：${result.sentences.size} 句"
+            }
+
+            override fun onError(error: Throwable) {
+                if (activeItemId != video.videoHash) {
+                    return
+                }
+                statusText.text = "句子边界加载失败：${error.message.orEmpty()}"
+                sentenceMetaText.text = "暂无句子边界"
+                sentenceText.text = video.title
+            }
+        })
+    }
+
+    private fun videoCard(
+        id: String,
+        title: String,
+        meta: String,
+        onClick: () -> Unit,
+    ): View {
+        val selected = activeItemId == id
+        val titleColor = if (selected) COLOR_TEXT_DARK else Color.WHITE
+        val metaColor = if (selected) COLOR_ACCENT_DEEP else COLOR_TEXT_MUTED
+        return LinearLayout(this).apply {
+            orientation = LinearLayout.VERTICAL
+            gravity = Gravity.CENTER_VERTICAL
+            isClickable = true
+            background = if (selected) {
+                rounded(COLOR_SELECTED, 8f, COLOR_ACCENT)
+            } else {
+                rounded(COLOR_SURFACE, 8f, COLOR_BORDER)
+            }
+            setPadding(14.dp, 12.dp, 14.dp, 12.dp)
+            layoutParams = LinearLayout.LayoutParams(218.dp, 84.dp).withMargins(0, 0, 10.dp, 0)
+            setOnClickListener { onClick() }
+            addView(TextView(this@MainActivity).apply {
+                text = title
+                setTextColor(titleColor)
+                textSize = 15f
+                typeface = Typeface.DEFAULT_BOLD
+                maxLines = 2
+            })
+            addView(TextView(this@MainActivity).apply {
+                text = meta
+                setTextColor(metaColor)
+                textSize = 12f
+                setPadding(0, 6.dp, 0, 0)
+            })
+        }
+    }
+
+    private fun headerButton(label: String, action: () -> Unit): Button {
         return Button(this).apply {
             text = label
-            setOnClickListener { action() }
+            isAllCaps = false
+            setTextColor(Color.WHITE)
+            textSize = 14f
+            typeface = Typeface.DEFAULT_BOLD
+            minWidth = 0
+            minHeight = 0
+            minimumWidth = 0
+            minimumHeight = 0
+            setPadding(14.dp, 0, 14.dp, 0)
+            background = rounded(COLOR_BUTTON, 8f)
             layoutParams = LinearLayout.LayoutParams(
-                0,
                 ViewGroup.LayoutParams.WRAP_CONTENT,
-                1f,
-            ).apply {
-                marginStart = 4
-                marginEnd = 4
-            }
+                40.dp,
+            ).withMargins(8.dp, 0, 0, 0)
+            setOnClickListener { action() }
+        }
+    }
+
+    private fun overlayPill(label: String): TextView {
+        return TextView(this).apply {
+            text = label
+            setTextColor(Color.WHITE)
+            textSize = 13f
+            typeface = Typeface.DEFAULT_BOLD
+            background = rounded(COLOR_OVERLAY, 8f)
+            setPadding(10.dp, 6.dp, 10.dp, 6.dp)
+            gravity = Gravity.CENTER
         }
     }
 
@@ -185,27 +518,103 @@ class MainActivity : Activity() {
 
     private fun sampleOnlineLesson(): TrainingMediaItem {
         return TrainingMediaItem(
-            id = "online_sample_01",
-            title = "Online Sample Lesson",
+            id = SAMPLE_ID,
+            title = "口语宝示例课",
             uri = Uri.parse("https://storage.googleapis.com/gtv-videos-bucket/sample/BigBuckBunny.mp4"),
             mimeType = MimeTypes.VIDEO_MP4,
-            customCacheKey = "online_sample_01",
+            customCacheKey = SAMPLE_ID,
             sentences = listOf(
-                SentenceBoundary(0, 0L, 3_000L, "Listen once, then imitate."),
-                SentenceBoundary(1, 3_000L, 7_000L, "Swipe left to go back."),
-                SentenceBoundary(2, 7_000L, 11_000L, "Swipe right to move forward."),
-                SentenceBoundary(3, 11_000L, 16_000L, "Tap the video area to pause or play."),
+                SentenceBoundary(0, 0L, 3_000L, "Good morning, everyone."),
+                SentenceBoundary(1, 3_000L, 7_000L, "Today we will practice listening carefully."),
+                SentenceBoundary(2, 7_000L, 11_000L, "Please repeat each sentence clearly."),
+                SentenceBoundary(3, 11_000L, 16_000L, "Small daily practice makes progress visible."),
             ),
         )
     }
 
-    private fun format(ms: Long): String {
+    private fun playbackStatus(snapshot: PlaybackSnapshot): String {
+        val state = when {
+            snapshot.isPlaying -> "播放中"
+            snapshot.playbackState.name == "BUFFERING" -> "缓冲中"
+            snapshot.playbackState.name == "ENDED" -> "已结束"
+            else -> "已暂停"
+        }
+        val sentence = snapshot.sentenceIndex?.let { " · 第 ${it + 1} 句" }.orEmpty()
+        return "$state$sentence"
+    }
+
+    private fun playbackProgress(snapshot: PlaybackSnapshot): Int {
+        if (snapshot.durationMs <= 0L) {
+            return 0
+        }
+        return ((snapshot.positionMs.coerceAtMost(snapshot.durationMs) * PROGRESS_MAX) / snapshot.durationMs).toInt()
+    }
+
+    private fun formatTime(ms: Long): String {
         if (ms < 0) {
             return "--:--"
         }
         val totalSeconds = ms / 1000
-        val minutes = totalSeconds / 60
+        val hours = totalSeconds / 3600
+        val minutes = (totalSeconds % 3600) / 60
         val seconds = totalSeconds % 60
-        return "%02d:%02d".format(minutes, seconds)
+        return if (hours > 0) {
+            "%d:%02d:%02d".format(Locale.US, hours, minutes, seconds)
+        } else {
+            "%02d:%02d".format(Locale.US, minutes, seconds)
+        }
+    }
+
+    private fun formatSpeed(speed: Float): String {
+        val value = String.format(Locale.US, "%.2f", speed).trimEnd('0').trimEnd('.')
+        return "${value}x"
+    }
+
+    private fun rounded(
+        color: Int,
+        radiusDp: Float,
+        strokeColor: Int? = null,
+    ): GradientDrawable {
+        return GradientDrawable().apply {
+            setColor(color)
+            cornerRadius = radiusDp.dp.toFloat()
+            strokeColor?.let { setStroke(1.dp, it) }
+        }
+    }
+
+    private fun LinearLayout.LayoutParams.withMargins(
+        left: Int,
+        top: Int,
+        right: Int,
+        bottom: Int,
+    ): LinearLayout.LayoutParams {
+        setMargins(left, top, right, bottom)
+        return this
+    }
+
+    private val Int.dp: Int
+        get() = (this * resources.displayMetrics.density).toInt()
+
+    private val Float.dp: Int
+        get() = (this * resources.displayMetrics.density).toInt()
+
+    private companion object {
+        const val PICK_VIDEO_REQUEST = 1001
+        const val SAMPLE_ID = "online_sample_01"
+        const val PROGRESS_MAX = 1000
+
+        val COLOR_BACKGROUND: Int = Color.rgb(12, 15, 18)
+        val COLOR_SURFACE: Int = Color.rgb(28, 34, 40)
+        val COLOR_BORDER: Int = Color.rgb(50, 59, 67)
+        val COLOR_BUTTON: Int = Color.rgb(37, 99, 235)
+        val COLOR_ACCENT: Int = Color.rgb(21, 184, 132)
+        val COLOR_ACCENT_DEEP: Int = Color.rgb(7, 118, 86)
+        val COLOR_SELECTED: Int = Color.rgb(229, 248, 240)
+        val COLOR_PROGRESS_TRACK: Int = Color.rgb(224, 231, 235)
+        val COLOR_LIGHT_BORDER: Int = Color.rgb(218, 226, 232)
+        val COLOR_OVERLAY: Int = Color.argb(178, 9, 12, 16)
+        val COLOR_TEXT_DARK: Int = Color.rgb(18, 24, 31)
+        val COLOR_TEXT_MUTED: Int = Color.rgb(151, 162, 174)
+        val COLOR_TEXT_SUBTLE: Int = Color.rgb(88, 98, 108)
     }
 }
diff --git a/android/sample-app/src/main/res/values/strings.xml b/android/sample-app/src/main/res/values/strings.xml
new file mode 100644
index 0000000..2aedfcc
--- /dev/null
+++ b/android/sample-app/src/main/res/values/strings.xml
@@ -0,0 +1,3 @@
+<resources>
+    <string name="app_name">口语宝</string>
+</resources>
diff --git a/output/karoline_audio.m4a b/output/karoline_audio.m4a
new file mode 100644
index 0000000..a16a811
Binary files /dev/null and b/output/karoline_audio.m4a differ
diff --git a/output/karoline_sample_30s.moss_0_9b_after_redownload.json b/output/karoline_sample_30s.moss_0_9b_after_redownload.json
new file mode 100644
index 0000000..8610ab7
--- /dev/null
+++ b/output/karoline_sample_30s.moss_0_9b_after_redownload.json
@@ -0,0 +1 @@
+{"error":{"message":"Model 'MOSS-Transcribe-Diarize-0.9B-MLX-5bit' failed to load: No safetensors found in /Users/liushuming/.omlx/models/aufklarer/MOSS-Transcribe-Diarize-0.9B-MLX-5bit. Reload models after fixing the files to retry.","type":"server_error","param":null,"code":null}}
\ No newline at end of file
diff --git a/output/karoline_sample_30s.transcription.json b/output/karoline_sample_30s.transcription.json
new file mode 100644
index 0000000..e97ea1a
--- /dev/null
+++ b/output/karoline_sample_30s.transcription.json
@@ -0,0 +1 @@
+{"error":{"message":"Model type moss_transcribe_diarize not supported.","type":"server_error","param":null,"code":null}}
\ No newline at end of file
diff --git a/output/karoline_sample_30s.wav b/output/karoline_sample_30s.wav
new file mode 100644
index 0000000..d4d67c3
Binary files /dev/null and b/output/karoline_sample_30s.wav differ
diff --git a/sentence_api/.env.example b/sentence_api/.env.example
new file mode 100644
index 0000000..d9af6fa
--- /dev/null
+++ b/sentence_api/.env.example
@@ -0,0 +1,14 @@
+ORAL_TRAINER_DATA_DIR=/app/sentence_api/data
+PUBLIC_BASE_URL=https://video_service.d1kt.cn
+ADMIN_API_KEY=replace-with-a-long-random-secret
+CLIENT_API_KEY=replace-with-a-separate-client-secret
+
+MAX_VIDEO_UPLOAD_BYTES=12884901888
+MAX_ATTEMPT_UPLOAD_BYTES=52428800
+KEEP_ATTEMPT_AUDIO=false
+ASSESSMENT_PASS_SCORE=70
+
+MOSS_TRANSCRIBE_URL=http://127.0.0.1:8001/v1/audio/transcriptions
+MOSS_MODEL=OpenMOSS-Team/MOSS-Transcribe-Diarize
+MOSS_TIMEOUT_SECONDS=1800
+MOSS_MAX_NEW_TOKENS=65536
diff --git a/sentence_api/DEPLOYMENT.md b/sentence_api/DEPLOYMENT.md
new file mode 100644
index 0000000..a8d1738
--- /dev/null
+++ b/sentence_api/DEPLOYMENT.md
@@ -0,0 +1,314 @@
+# 口语训练视频服务部署指南
+
+本文部署以下两个进程：
+
+- `oral-trainer-api`：视频上传、管理后台、边界查询、Range 视频播放和朗读评分。
+- `MOSS-Transcribe-Diarize`：GPU 转写服务，只在服务器内网监听。
+
+推荐 Ubuntu 22.04/24.04、Python 3.12、FFmpeg、NVIDIA GPU 和已安装的 NVIDIA 驱动。MOSS 0.9B 的实际显存占用受推理框架、并发和音频长度影响，生产环境建议从 16 GB 以上显存开始验证。
+
+## 1. 目录与端口
+
+本文假设代码位于 `/opt/oral-trainer`：
+
+```text
+/opt/oral-trainer/
+  sentence_api/
+    data/
+      v/                 上传的视频
+      work/              临时转码文件
+      attempts/          可选的学生录音
+      oral_trainer.sqlite3
+```
+
+端口规划：
+
+| 服务 | 监听地址 | 用途 |
+| --- | --- | --- |
+| MOSS | `127.0.0.1:8001` | 内部语音转写 |
+| FastAPI | `127.0.0.1:8000` | 内部应用服务 |
+| Nginx | `0.0.0.0:443` | 对外 HTTPS |
+
+不要把 MOSS 的 `8001` 端口直接暴露到公网。
+
+## 2. 部署 MOSS
+
+官方模型为 `OpenMOSS-Team/MOSS-Transcribe-Diarize`。CUDA 12 环境可使用官方当前说明中的 vLLM 构建：
+
+```bash
+sudo apt update
+sudo apt install -y ffmpeg git curl
+curl -LsSf https://astral.sh/uv/install.sh | sh
+
+sudo mkdir -p /opt/moss-transcribe
+sudo chown "$USER":"$USER" /opt/moss-transcribe
+cd /opt/moss-transcribe
+
+uv venv --python 3.12 .venv
+. .venv/bin/activate
+uv pip install -U vllm \
+  --torch-backend=auto \
+  --extra-index-url https://wheels.vllm.ai/68b4a1d582818e67adc903bf1b8fc5a5447da2fa/cu129
+```
+
+启动模型：
+
+```bash
+. /opt/moss-transcribe/.venv/bin/activate
+vllm serve OpenMOSS-Team/MOSS-Transcribe-Diarize \
+  --host 127.0.0.1 \
+  --port 8001 \
+  --trust-remote-code
+```
+
+生产环境可把它注册为 systemd 服务：
+
+```ini
+# /etc/systemd/system/moss-transcribe.service
+[Unit]
+Description=MOSS Transcribe Service
+After=network-online.target
+
+[Service]
+Type=simple
+User=oraltrainer
+WorkingDirectory=/opt/moss-transcribe
+ExecStart=/opt/moss-transcribe/.venv/bin/vllm serve OpenMOSS-Team/MOSS-Transcribe-Diarize --host 127.0.0.1 --port 8001 --trust-remote-code
+Restart=on-failure
+RestartSec=10
+TimeoutStartSec=1800
+
+[Install]
+WantedBy=multi-user.target
+```
+
+```bash
+sudo systemctl daemon-reload
+sudo systemctl enable --now moss-transcribe
+sudo journalctl -u moss-transcribe -f
+```
+
+模型第一次启动会下载权重。CUDA 13 服务器也可以按照 MOSS 官方仓库说明改用 SGLang Omni。推理框架的安装地址可能随上游版本更新，正式部署前应对照 [MOSS 官方仓库](https://github.com/OpenMOSS/MOSS-Transcribe-Diarize) 的 Quickstart。
+
+验证：
+
+```bash
+curl -X POST http://127.0.0.1:8001/v1/audio/transcriptions \
+  -F model=OpenMOSS-Team/MOSS-Transcribe-Diarize \
+  -F file=@/path/to/test.wav \
+  -F response_format=verbose_json \
+  -F temperature=0
+```
+
+响应应包含 `text`；开启 `verbose_json` 后应包含带 `start`、`end`、`text` 的 `segments`。
+
+## 3. 部署 API
+
+### Docker Compose 方式
+
+```bash
+cd /opt/oral-trainer/sentence_api
+cp .env.example .env
+openssl rand -hex 32
+```
+
+分别生成两个随机值，填入 `.env` 的 `ADMIN_API_KEY` 和 `CLIENT_API_KEY`，并确认：
+
+```dotenv
+PUBLIC_BASE_URL=https://video_service.d1kt.cn
+MOSS_TRANSCRIBE_URL=http://127.0.0.1:8001/v1/audio/transcriptions
+```
+
+启动：
+
+```bash
+docker compose up -d --build
+docker compose logs -f oral-trainer-api
+```
+
+Compose 配置使用 Linux 的 host network，使容器可以访问只监听
+`127.0.0.1:8001` 的 MOSS，同时 API 也只监听 `127.0.0.1:8000`。如果在
+Docker Desktop 上做本地测试，可移除 `network_mode: host`，恢复端口映射，并把
+MOSS 地址改为 `host.docker.internal`。
+
+检查：
+
+```bash
+curl http://127.0.0.1:8000/healthz
+```
+
+`moss_configured` 应为 `true`。这里只表示地址已配置，实际连通性会在上传视频或评分时验证。
+
+### 不使用 Docker
+
+```bash
+cd /opt/oral-trainer
+python3.12 -m venv .venv-api
+. .venv-api/bin/activate
+python -m pip install -r sentence_api/requirements.txt
+
+export ORAL_TRAINER_DATA_DIR=/opt/oral-trainer-data
+export PUBLIC_BASE_URL=https://video_service.d1kt.cn
+export ADMIN_API_KEY='替换为随机密钥'
+export MOSS_TRANSCRIBE_URL=http://127.0.0.1:8001/v1/audio/transcriptions
+
+python -m uvicorn sentence_api.main:app \
+  --host 127.0.0.1 --port 8000 --workers 1 \
+  --proxy-headers --forwarded-allow-ips='*'
+```
+
+不使用 Docker 时可创建 API 的 systemd 服务：
+
+```ini
+# /etc/systemd/system/oral-trainer-api.service
+[Unit]
+Description=Oral Trainer Video API
+After=network-online.target moss-transcribe.service
+
+[Service]
+Type=simple
+User=oraltrainer
+WorkingDirectory=/opt/oral-trainer
+EnvironmentFile=/opt/oral-trainer/sentence_api/.env
+ExecStart=/opt/oral-trainer/.venv-api/bin/python -m uvicorn sentence_api.main:app --host 127.0.0.1 --port 8000 --workers 1 --proxy-headers --forwarded-allow-ips=*
+Restart=on-failure
+RestartSec=5
+
+[Install]
+WantedBy=multi-user.target
+```
+
+```bash
+sudo systemctl daemon-reload
+sudo systemctl enable --now oral-trainer-api
+sudo journalctl -u oral-trainer-api -f
+```
+
+当前上传后的处理任务运行在 API 进程中，因此只能使用一个 Uvicorn worker。需要多机或多 worker 时，应把 `VideoProcessor.process` 迁移到 Celery、RQ 或其他持久化任务队列。
+
+## 4. 配置 Nginx 与 HTTPS
+
+```bash
+sudo cp sentence_api/nginx.conf.example /etc/nginx/sites-available/oral-trainer
+sudo ln -s /etc/nginx/sites-available/oral-trainer /etc/nginx/sites-enabled/oral-trainer
+sudo nginx -t
+sudo systemctl reload nginx
+```
+
+配置文件中的关键项：
+
+- `client_max_body_size 12g`：允许超过 1 GB 的视频。
+- `proxy_request_buffering off`：上传时直接流向 FastAPI，避免 Nginx 再完整缓存一份。
+- `proxy_force_ranges on`：保留 Android 随机拖动播放所需的 HTTP Range。
+- `proxy_read_timeout 7200s`：允许长视频处理和慢速上传。
+
+签发证书：
+
+```bash
+sudo apt install -y certbot python3-certbot-nginx
+sudo certbot --nginx -d video_service.d1kt.cn
+```
+
+证书完成后验证：
+
+```bash
+curl https://video_service.d1kt.cn/healthz
+```
+
+管理后台地址：
+
+```text
+https://video_service.d1kt.cn/admin
+```
+
+后台的管理密钥保存在当前浏览器的 `localStorage`，接口请求通过 `X-Admin-Key` 发送。
+`CLIENT_API_KEY` 应通过 Android 构建配置注入 `OralTrainerSdkConfig.assessmentApiKey`，
+不要硬编码进公开代码仓库。SDK 会在评分请求中发送 `X-Client-Key`。公开视频列表和播放
+接口仍可交给 CDN 缓存，GPU 评分接口则受到密钥保护。
+
+## 5. 上传与处理流程
+
+后台上传后，服务会：
+
+1. 管理后台通过 raw body 流式写入临时文件并同步计算 SHA-256，不会先在系统临时目录保留完整副本。兼容的 multipart 接口仍以 4 MB 分块复制。
+2. 保存到 `data/v/{sha256}.{extension}`。
+3. 使用 FFmpeg 提取 16 kHz 单声道 WAV。
+4. 请求 MOSS 的 `verbose_json` 转写接口。
+5. 以 MOSS 时间戳生成句子边界，并计算每句原音有效语音时长。
+6. 把视频和句子写入 SQLite，状态变为 `ready`。
+
+查询处理状态：
+
+```bash
+curl https://video_service.d1kt.cn/api/v1/videos
+curl https://video_service.d1kt.cn/api/v1/videos/{sha256}
+```
+
+如果 MOSS 未配置，上传仍会使用原来的静音检测生成边界，但句子文本为空，无法开始测验。配置好 MOSS 后可在后台点击“重新处理”。
+
+## 6. 视频随机播放
+
+移动端播放地址为：
+
+```text
+GET /api/v1/videos/{sha256}/content
+```
+
+服务支持 HTTP Range，Media3/ExoPlayer 可以随机 seek，并继续使用 SDK 现有的本地缓存。第一阶段不必强制改成 HLS。
+
+上传前建议把 MP4 处理成 H.264/AAC 并把 `moov` 元数据移动到文件头：
+
+```bash
+ffmpeg -i input.mp4 -c copy -movflags +faststart output.mp4
+```
+
+如果原始编码不被 Android 广泛支持，再转码：
+
+```bash
+ffmpeg -i input.mkv \
+  -c:v libx264 -preset medium -crf 22 \
+  -c:a aac -b:a 128k -movflags +faststart output.mp4
+```
+
+当并发量明显增长或网络波动较大时，再增加 HLS/DASH 转码和对象存储/CDN。
+
+## 7. 朗读评分接口
+
+```bash
+curl -X POST \
+  https://video_service.d1kt.cn/api/v1/videos/{sha256}/sentences/0/assessments \
+  -H 'X-Client-Key: 替换为客户端密钥' \
+  -F audio=@student.wav \
+  -F language=en
+```
+
+当前 `asr-fluency-v1` 评分为：
+
+```text
+总分 = 内容分 * 80% + 流畅度 * 20%
+流畅度 = 有效语音时长分 * 35% + 停顿分 * 40% + 语速分 * 25%
+```
+
+总分不低于 `ASSESSMENT_PASS_SCORE`（默认 70）即通过。`pronunciation_score` 和 `prosody_score` 目前返回 `null`，避免把 ASR 内容匹配误报为音素发音质量。
+
+服务只向 MOSS 发送语言提示，不发送标准句子作为 prompt 或热词，避免标准答案诱导识别结果。
+
+## 8. 运维与备份
+
+需要持久化备份：
+
+```text
+sentence_api/data/v/
+sentence_api/data/oral_trainer.sqlite3
+```
+
+`work/` 可以清理。`KEEP_ATTEMPT_AUDIO=false` 时学生录音在评分完成后自动删除；设为 `true` 时还需备份和定期清理 `attempts/`，并在产品隐私政策中说明保存期限。
+
+SQLite 在线备份示例：
+
+```bash
+sqlite3 sentence_api/data/oral_trainer.sqlite3 \
+  ".backup '/backup/oral_trainer-$(date +%F).sqlite3'"
+```
+
+升级前先备份数据库和 `data/v/`。API 重启期间被中断的视频会保留 `processing` 状态，可从后台执行“重新处理”。
diff --git a/sentence_api/Dockerfile b/sentence_api/Dockerfile
new file mode 100644
index 0000000..a01d57a
--- /dev/null
+++ b/sentence_api/Dockerfile
@@ -0,0 +1,20 @@
+FROM python:3.12-slim
+
+ENV PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1 \
+    PIP_NO_CACHE_DIR=1
+
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends ffmpeg ca-certificates \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+COPY sentence_api/requirements.txt /app/sentence_api/requirements.txt
+RUN python -m pip install --upgrade pip \
+    && python -m pip install -r /app/sentence_api/requirements.txt
+
+COPY sentence_api /app/sentence_api
+COPY sentence_analysis.py /app/sentence_analysis.py
+
+EXPOSE 8000
+CMD ["python", "-m", "uvicorn", "sentence_api.main:app", "--host", "0.0.0.0", "--port", "8000", "--workers", "1", "--proxy-headers", "--forwarded-allow-ips", "*"]
diff --git a/sentence_api/README.md b/sentence_api/README.md
index afbfc60..af76507 100644
--- a/sentence_api/README.md
+++ b/sentence_api/README.md
@@ -1,92 +1,82 @@
-# Sentence Boundary API
+# Oral Trainer Video Service
 
-This service looks up pre-generated sentence boundaries by the SHA-256 hash of
-the exact video bytes. It does not analyze media during an API request.
+FastAPI service for the Android oral-training SDK. It provides:
 
-## Install
+- chunked video upload with SHA-256 calculation;
+- files stored under `data/v/` and metadata stored in SQLite;
+- background sentence transcription through MOSS-Transcribe-Diarize;
+- a browser-based administration page at `/admin`;
+- video catalog and HTTP Range playback;
+- sentence-boundary lookup compatible with the existing Android SDK;
+- student recording upload and `asr-fluency-v1` assessment;
+- the legacy JSON boundary index as a read-only fallback.
 
-From the repository root:
+## Local Run
 
 ```bash
-python3 -m venv .venv-sentence-api
+python3.12 -m venv .venv-sentence-api
 . .venv-sentence-api/bin/activate
 python -m pip install -r sentence_api/requirements.txt
-```
 
-## Generate The Index
-
-Generate boundaries with the same silence detector used by the desktop player:
-
-```bash
-python -m sentence_api.generate_boundaries \
-  "/path/to/lesson.mp4" \
-  --index sentence_api/data/sentence_boundaries.json
-```
-
-The command calculates the SHA-256 hash, detects boundaries, converts seconds
-to milliseconds, infers each `end_ms` from the next sentence start, and writes
-the result atomically into the JSON index. The last sentence ends at the media
-duration.
-
-For a large course library, run this command in an ingestion worker and store
-the same document in a database or object storage instead of committing the
-JSON file to the application image.
-
-## Run
-
-```bash
-SENTENCE_BOUNDARIES_FILE=sentence_api/data/sentence_boundaries.json \
+export ADMIN_API_KEY=local-development-key
+export MOSS_TRANSCRIBE_URL=http://127.0.0.1:8001/v1/audio/transcriptions
 python -m uvicorn sentence_api.main:app --host 0.0.0.0 --port 8000
 ```
 
-The interactive API documentation is available at `/docs`.
+Open:
 
-## Request
-
-```http
-GET /api/v1/videos/{sha256}/sentence-boundaries
+```text
+http://127.0.0.1:8000/admin
+http://127.0.0.1:8000/docs
 ```
 
-Example using the demo record in the checked-in index:
+When `MOSS_TRANSCRIBE_URL` is empty, uploaded videos still receive silence-based
+boundaries, but no reference transcript is produced and assessment is disabled.
+
+## Main Endpoints
+
+```text
+GET    /healthz
+GET    /api/v1/videos
+GET    /api/v1/videos/{sha256}
+GET    /api/v1/videos/{sha256}/content
+GET    /api/v1/videos/{sha256}/sentence-boundaries
+POST   /api/v1/videos/{sha256}/sentences/{index}/assessments
+
+POST   /api/v1/admin/videos
+PUT    /api/v1/admin/videos/raw
+POST   /api/v1/admin/videos/{sha256}/process
+PUT    /api/v1/admin/videos/{sha256}/sentences/{index}
+DELETE /api/v1/admin/videos/{sha256}
+```
+
+Admin endpoints use `X-Admin-Key` when `ADMIN_API_KEY` is configured. Assessment
+requests use `X-Client-Key` when `CLIENT_API_KEY` is configured.
+
+## Assessment
 
 ```bash
-curl http://127.0.0.1:8000/api/v1/videos/468a4d064f6ec49942b45e25ab93c500d31870f978c4f28ff8b3b408852326e0/sentence-boundaries
+curl -X POST \
+  http://127.0.0.1:8000/api/v1/videos/{sha256}/sentences/0/assessments \
+  -F audio=@student.wav \
+  -F language=en
 ```
 
-The MP4 used during development is also indexed. Its hash is
-`b6631d5cf48f37fed0ecc623563dd48b7ed660689b4d25d2ebac7fecab807ddc`, and its
-generated index contains 244 boundaries.
+The current score intentionally measures reading content and fluency:
 
-The response is:
-
-```json
-{
-  "video_hash": "...",
-  "duration_ms": 16000,
-  "algorithm_version": "silence-rms-v1",
-  "sentences": [
-    {
-      "index": 0,
-      "start_ms": 0,
-      "end_ms": 4230,
-      "text": null
-    }
-  ]
-}
+```text
+overall = content * 80% + fluency * 20%
+fluency = duration * 35% + pauses * 40% + speech rate * 25%
 ```
 
-Unknown hashes return `404`. A hash must be a 64-character hexadecimal
-SHA-256 digest; malformed values return `422`.
+`pronunciation_score` and `prosody_score` remain `null` until a phoneme/GOP
+model is connected.
 
-When testing from a physical Android phone, replace `127.0.0.1` with the
-computer's LAN IP address. `127.0.0.1` on the phone refers to the phone itself.
-For production, expose the API over HTTPS.
+## Tests
 
-## Android Request Flow
+```bash
+python -m pytest sentence_api/tests -q
+```
 
-The mobile app should calculate the hash from the selected `content://` URI in
-streaming chunks, request the endpoint, map the returned `sentences` to
-`SentenceBoundary`, and then call `controller.loadItem`. The hash must be
-calculated from the exact bytes of the same video served to the player. For
-HTTPS course videos, the course manifest can carry the hash and avoid hashing
-the entire remote file on every device.
+See [DEPLOYMENT.md](DEPLOYMENT.md) for Docker, MOSS, Nginx, HTTPS, large-file
+upload, backup, and production operation instructions.
diff --git a/sentence_api/assessment.py b/sentence_api/assessment.py
new file mode 100644
index 0000000..2bf3f88
--- /dev/null
+++ b/sentence_api/assessment.py
@@ -0,0 +1,107 @@
+import uuid
+from pathlib import Path
+from typing import Optional
+
+from .audio_metrics import analyze_audio
+from .models import AssessmentResult, SentenceBoundaryDocument, TextSubstitution
+from .repository import VideoRepository
+from .scoring import SCORING_VERSION, score_attempt
+from .transcription import Transcriber
+
+
+class AssessmentService:
+    def __init__(
+        self,
+        repository: VideoRepository,
+        transcriber: Transcriber,
+        pass_score: float,
+    ):
+        self.repository = repository
+        self.transcriber = transcriber
+        self.pass_score = pass_score
+
+    def assess(
+        self,
+        *,
+        document: SentenceBoundaryDocument,
+        sentence_index: int,
+        audio_path: Path,
+        language: Optional[str] = None,
+        retained_audio_filename: Optional[str] = None,
+    ) -> AssessmentResult:
+        if not self.transcriber.available:
+            raise RuntimeError("MOSS transcription is not configured on this server.")
+        sentence = next(
+            (item for item in document.sentences if item.index == sentence_index),
+            None,
+        )
+        if sentence is None:
+            raise LookupError("Sentence index was not found.")
+        if not sentence.text or not sentence.text.strip():
+            raise ValueError("The sentence has no reviewed reference text yet.")
+
+        student_metrics = analyze_audio(audio_path)
+        transcript = self.transcriber.transcribe(
+            audio_path,
+            language or sentence.language,
+        )
+        recognized_text = transcript.text.strip()
+        if not recognized_text:
+            raise ValueError("MOSS did not recognize any speech in the recording.")
+
+        reference_duration_ms = sentence.end_ms - sentence.start_ms
+        reference_speech_duration_ms = (
+            sentence.reference_speech_duration_ms or reference_duration_ms
+        )
+        score = score_attempt(
+            reference_text=sentence.text,
+            recognized_text=recognized_text,
+            reference_speech_duration_ms=reference_speech_duration_ms,
+            student_metrics=student_metrics,
+        )
+        attempt_id = uuid.uuid4().hex
+        result = AssessmentResult(
+            attempt_id=attempt_id,
+            scoring_version=SCORING_VERSION,
+            overall_score=score.overall_score,
+            passed=score.overall_score >= self.pass_score,
+            pass_score=self.pass_score,
+            content_score=score.content_score,
+            completeness_score=score.completeness_score,
+            fluency_score=score.fluency_score,
+            pronunciation_score=None,
+            prosody_score=None,
+            duration_score=score.duration_score,
+            pause_score=score.pause_score,
+            speech_rate_score=score.speech_rate_score,
+            reference_text=sentence.text,
+            recognized_text=recognized_text,
+            reference_duration_ms=reference_duration_ms,
+            reference_speech_duration_ms=reference_speech_duration_ms,
+            student_recording_duration_ms=student_metrics.recording_duration_ms,
+            student_speech_duration_ms=student_metrics.speech_duration_ms,
+            duration_ratio=score.duration_ratio,
+            missing_tokens=score.missing_tokens,
+            extra_tokens=score.extra_tokens,
+            substitutions=[
+                TextSubstitution(expected=expected, actual=actual)
+                for expected, actual in score.substitutions
+            ],
+            feedback=score.feedback,
+            details={
+                "score_kind": "朗读内容与流畅度匹配分",
+                "content_weight": "0.80",
+                "fluency_weight": "0.20",
+                "duration_weight_within_fluency": "0.35",
+                "phoneme_scoring": "not_enabled",
+            },
+        )
+        if self.repository.get_video(document.video_hash) is not None:
+            self.repository.record_attempt(
+                attempt_id=attempt_id,
+                video_hash=document.video_hash,
+                sentence_index=sentence_index,
+                result=result.model_dump(mode="json"),
+                audio_filename=retained_audio_filename,
+            )
+        return result
diff --git a/sentence_api/audio_metrics.py b/sentence_api/audio_metrics.py
new file mode 100644
index 0000000..a9b65ca
--- /dev/null
+++ b/sentence_api/audio_metrics.py
@@ -0,0 +1,123 @@
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Tuple
+
+import numpy as np
+
+
+SAMPLE_RATE = 16_000
+FRAME_SAMPLES = 480
+
+
+class AudioAnalysisError(RuntimeError):
+    pass
+
+
+@dataclass(frozen=True)
+class AudioMetrics:
+    recording_duration_ms: int
+    speech_duration_ms: int
+    internal_silence_ms: int
+    internal_pause_ratio: float
+
+
+def decode_audio_mono(path: Path) -> Tuple[np.ndarray, int]:
+    try:
+        import av
+    except ImportError as exc:
+        raise AudioAnalysisError("PyAV is required for audio analysis.") from exc
+
+    container = av.open(str(path))
+    chunks = []
+    try:
+        audio_stream = next(
+            (stream for stream in container.streams if stream.type == "audio"),
+            None,
+        )
+        if audio_stream is None:
+            raise AudioAnalysisError("The uploaded file does not contain an audio stream.")
+        resampler = av.AudioResampler(format="fltp", layout="mono", rate=SAMPLE_RATE)
+        for packet in container.demux(audio_stream):
+            for frame in packet.decode():
+                for output in resampler.resample(frame):
+                    chunks.append(output.to_ndarray()[0].astype(np.float32, copy=False))
+        for output in resampler.resample(None):
+            chunks.append(output.to_ndarray()[0].astype(np.float32, copy=False))
+    finally:
+        container.close()
+
+    if not chunks:
+        raise AudioAnalysisError("The uploaded audio is empty.")
+    return np.concatenate(chunks), SAMPLE_RATE
+
+
+def analyze_audio(path: Path) -> AudioMetrics:
+    samples, sample_rate = decode_audio_mono(path)
+    return analyze_samples(samples, sample_rate)
+
+
+def analyze_samples(samples: np.ndarray, sample_rate: int = SAMPLE_RATE) -> AudioMetrics:
+    if samples.ndim != 1:
+        samples = samples.reshape(-1)
+    if samples.size == 0 or sample_rate <= 0:
+        raise AudioAnalysisError("The uploaded audio is empty.")
+
+    recording_duration_ms = max(1, int(round(samples.size / sample_rate * 1000)))
+    window = max(1, int(round(sample_rate * 0.03)))
+    complete_frames = int(np.ceil(samples.size / window))
+    padded = np.pad(samples, (0, complete_frames * window - samples.size))
+    frames = padded.reshape(complete_frames, window).astype(np.float64, copy=False)
+    rms = np.sqrt(np.mean(frames * frames, axis=1))
+
+    signal_level = float(np.percentile(rms, 95))
+    if signal_level < 0.002:
+        raise AudioAnalysisError("No usable speech was detected in the recording.")
+    noise_floor = float(np.percentile(rms, 10))
+    threshold = min(signal_level * 0.45, max(0.006, noise_floor * 2.2))
+    speech = rms >= threshold
+
+    # Treat very short gaps inside a word as speech, then reject short clicks.
+    _bridge_false_runs(speech, max_frames=5)
+    _remove_true_runs(speech, max_frames=2)
+    speech_indexes = np.flatnonzero(speech)
+    if speech_indexes.size == 0:
+        raise AudioAnalysisError("No usable speech was detected in the recording.")
+
+    frame_ms = window / sample_rate * 1000
+    speech_duration_ms = max(1, int(round(speech.sum() * frame_ms)))
+    first = int(speech_indexes[0])
+    last = int(speech_indexes[-1])
+    internal_frames = max(1, last - first + 1)
+    internal_silence_frames = int((~speech[first : last + 1]).sum())
+    internal_silence_ms = int(round(internal_silence_frames * frame_ms))
+    internal_pause_ratio = internal_silence_frames / internal_frames
+    return AudioMetrics(
+        recording_duration_ms=recording_duration_ms,
+        speech_duration_ms=speech_duration_ms,
+        internal_silence_ms=internal_silence_ms,
+        internal_pause_ratio=round(internal_pause_ratio, 4),
+    )
+
+
+def _bridge_false_runs(values: np.ndarray, max_frames: int) -> None:
+    start = None
+    for index, value in enumerate(values):
+        if not value and start is None:
+            start = index
+        elif value and start is not None:
+            if start > 0 and index - start <= max_frames:
+                values[start:index] = True
+            start = None
+
+
+def _remove_true_runs(values: np.ndarray, max_frames: int) -> None:
+    start = None
+    for index, value in enumerate(values):
+        if value and start is None:
+            start = index
+        elif not value and start is not None:
+            if index - start <= max_frames:
+                values[start:index] = False
+            start = None
+    if start is not None and len(values) - start <= max_frames:
+        values[start:] = False
diff --git a/sentence_api/config.py b/sentence_api/config.py
new file mode 100644
index 0000000..6544424
--- /dev/null
+++ b/sentence_api/config.py
@@ -0,0 +1,75 @@
+import os
+from dataclasses import dataclass
+from pathlib import Path
+
+
+def _bool_env(name: str, default: bool) -> bool:
+    value = os.getenv(name)
+    if value is None:
+        return default
+    return value.strip().lower() in {"1", "true", "yes", "on"}
+
+
+@dataclass(frozen=True)
+class Settings:
+    data_dir: Path
+    legacy_boundaries_file: Path
+    admin_api_key: str
+    client_api_key: str
+    public_base_url: str
+    max_upload_bytes: int
+    max_attempt_bytes: int
+    keep_attempt_audio: bool
+    moss_transcribe_url: str
+    moss_model: str
+    moss_timeout_seconds: float
+    moss_max_new_tokens: int
+    pass_score: float
+
+    @classmethod
+    def from_env(cls) -> "Settings":
+        package_dir = Path(__file__).resolve().parent
+        data_dir = Path(os.getenv("ORAL_TRAINER_DATA_DIR", str(package_dir / "data")))
+        return cls(
+            data_dir=data_dir,
+            legacy_boundaries_file=Path(
+                os.getenv(
+                    "SENTENCE_BOUNDARIES_FILE",
+                    str(package_dir / "data" / "sentence_boundaries.json"),
+                )
+            ),
+            admin_api_key=os.getenv("ADMIN_API_KEY", ""),
+            client_api_key=os.getenv("CLIENT_API_KEY", ""),
+            public_base_url=os.getenv("PUBLIC_BASE_URL", "").rstrip("/"),
+            max_upload_bytes=int(os.getenv("MAX_VIDEO_UPLOAD_BYTES", str(12 * 1024**3))),
+            max_attempt_bytes=int(os.getenv("MAX_ATTEMPT_UPLOAD_BYTES", str(50 * 1024**2))),
+            keep_attempt_audio=_bool_env("KEEP_ATTEMPT_AUDIO", False),
+            moss_transcribe_url=os.getenv("MOSS_TRANSCRIBE_URL", "").strip(),
+            moss_model=os.getenv(
+                "MOSS_MODEL",
+                "OpenMOSS-Team/MOSS-Transcribe-Diarize",
+            ),
+            moss_timeout_seconds=float(os.getenv("MOSS_TIMEOUT_SECONDS", "1800")),
+            moss_max_new_tokens=int(os.getenv("MOSS_MAX_NEW_TOKENS", "65536")),
+            pass_score=float(os.getenv("ASSESSMENT_PASS_SCORE", "70")),
+        )
+
+    @property
+    def videos_dir(self) -> Path:
+        return self.data_dir / "v"
+
+    @property
+    def work_dir(self) -> Path:
+        return self.data_dir / "work"
+
+    @property
+    def attempts_dir(self) -> Path:
+        return self.data_dir / "attempts"
+
+    @property
+    def database_path(self) -> Path:
+        return self.data_dir / "oral_trainer.sqlite3"
+
+    def ensure_directories(self) -> None:
+        for path in (self.data_dir, self.videos_dir, self.work_dir, self.attempts_dir):
+            path.mkdir(parents=True, exist_ok=True)
diff --git a/sentence_api/docker-compose.yml b/sentence_api/docker-compose.yml
new file mode 100644
index 0000000..aef7d55
--- /dev/null
+++ b/sentence_api/docker-compose.yml
@@ -0,0 +1,17 @@
+services:
+  oral-trainer-api:
+    build:
+      context: ..
+      dockerfile: sentence_api/Dockerfile
+    restart: unless-stopped
+    network_mode: host
+    env_file:
+      - .env
+    command: ["python", "-m", "uvicorn", "sentence_api.main:app", "--host", "127.0.0.1", "--port", "8000", "--workers", "1", "--proxy-headers", "--forwarded-allow-ips", "*"]
+    volumes:
+      - ./data:/app/sentence_api/data
+    healthcheck:
+      test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8000/healthz', timeout=5)"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
diff --git a/sentence_api/main.py b/sentence_api/main.py
index f139172..8899f2a 100644
--- a/sentence_api/main.py
+++ b/sentence_api/main.py
@@ -1,31 +1,161 @@
-import os
+import hashlib
+import hmac
+import logging
+import mimetypes
+import uuid
 from pathlib import Path
-from typing import Any, Dict
+from typing import Any, BinaryIO, Dict, Optional
 
-from fastapi import Depends, FastAPI, HTTPException, Path as ApiPath
+from fastapi import (
+    BackgroundTasks,
+    Depends,
+    FastAPI,
+    File,
+    Form,
+    Header,
+    HTTPException,
+    Path as ApiPath,
+    Query,
+    Request,
+    UploadFile,
+)
+from fastapi.responses import FileResponse, Response
+from fastapi.staticfiles import StaticFiles
+from starlette.concurrency import run_in_threadpool
 
-from .models import SentenceBoundaryDocument
+from .assessment import AssessmentService
+from .audio_metrics import AudioAnalysisError
+from .config import Settings
+from .models import (
+    AssessmentResult,
+    SentenceBoundary,
+    SentenceBoundaryDocument,
+    SentenceTextUpdate,
+    VideoDetailResponse,
+    VideoListResponse,
+    VideoSummary,
+    VideoUploadResponse,
+)
+from .processing import VideoProcessor
+from .repository import VideoRepository
 from .store import BoundaryStore
+from .transcription import MossTranscriber, Transcriber
 
 
-DEFAULT_INDEX_PATH = Path(__file__).resolve().parent / "data" / "sentence_boundaries.json"
+logger = logging.getLogger(__name__)
+SHA256_PATH = ApiPath(
+    min_length=64,
+    max_length=64,
+    pattern=r"^[A-Fa-f0-9]{64}$",
+    description="SHA-256 hex digest of the exact uploaded video bytes",
+)
+ALLOWED_VIDEO_EXTENSIONS = {".mp4", ".m4v", ".mov", ".mkv", ".webm"}
 
 
-def create_app(store: BoundaryStore = None) -> FastAPI:
-    application = FastAPI(
-        title="Oral Trainer Sentence Boundary API",
-        version="1.0.0",
-        description="Looks up pre-generated sentence boundaries by video SHA-256.",
+class UploadTooLargeError(ValueError):
+    pass
+
+
+def create_app(
+    store: Optional[BoundaryStore] = None,
+    *,
+    settings: Optional[Settings] = None,
+    repository: Optional[VideoRepository] = None,
+    transcriber: Optional[Transcriber] = None,
+) -> FastAPI:
+    service_settings = settings or Settings.from_env()
+    service_settings.ensure_directories()
+    video_repository = repository or VideoRepository(service_settings.database_path)
+    legacy_store = store or BoundaryStore(service_settings.legacy_boundaries_file)
+    moss = transcriber or MossTranscriber(
+        endpoint=service_settings.moss_transcribe_url,
+        model=service_settings.moss_model,
+        timeout_seconds=service_settings.moss_timeout_seconds,
+        max_new_tokens=service_settings.moss_max_new_tokens,
+    )
+    processor = VideoProcessor(service_settings, video_repository, moss)
+    assessment_service = AssessmentService(
+        video_repository,
+        moss,
+        service_settings.pass_score,
     )
-    index_path = Path(os.getenv("SENTENCE_BOUNDARIES_FILE", str(DEFAULT_INDEX_PATH)))
-    application.state.boundary_store = store or BoundaryStore(index_path)
 
-    def get_store() -> BoundaryStore:
-        return application.state.boundary_store
+    application = FastAPI(
+        title="Oral Trainer Video Service",
+        version="2.0.0",
+        description="Video ingestion, sentence transcription, streaming, and oral-reading assessment.",
+    )
+    application.state.settings = service_settings
+    application.state.boundary_store = legacy_store
+    application.state.video_repository = video_repository
+    application.state.transcriber = moss
+    application.state.video_processor = processor
+    application.state.assessment_service = assessment_service
+
+    static_dir = Path(__file__).resolve().parent / "static"
+    if static_dir.is_dir():
+        application.mount("/static", StaticFiles(directory=static_dir), name="static")
+
+    def require_admin(x_admin_key: Optional[str] = Header(default=None)) -> None:
+        expected = service_settings.admin_api_key
+        if expected and not hmac.compare_digest(x_admin_key or "", expected):
+            raise HTTPException(status_code=401, detail="A valid X-Admin-Key header is required.")
+
+    def require_client(x_client_key: Optional[str] = Header(default=None)) -> None:
+        expected = service_settings.client_api_key
+        if expected and not hmac.compare_digest(x_client_key or "", expected):
+            raise HTTPException(status_code=401, detail="A valid X-Client-Key header is required.")
+
+    def find_document(video_hash: str) -> Optional[SentenceBoundaryDocument]:
+        return video_repository.get_document(video_hash) or legacy_store.get(video_hash)
 
     @application.get("/healthz")
-    def healthz(boundary_store: BoundaryStore = Depends(get_store)) -> Dict[str, Any]:
-        return {"status": "ok", "video_count": boundary_store.count()}
+    def healthz() -> Dict[str, Any]:
+        videos = video_repository.list_videos()
+        return {
+            "status": "ok",
+            "video_count": len(videos) + legacy_store.count(),
+            "managed_video_count": len(videos),
+            "moss_configured": moss.available,
+            "scoring_version": "asr-fluency-v1",
+        }
+
+    @application.get("/admin", include_in_schema=False)
+    def admin_page() -> FileResponse:
+        page = static_dir / "admin.html"
+        if not page.is_file():
+            raise HTTPException(status_code=404, detail="Admin UI is not installed.")
+        return FileResponse(page)
+
+    @application.get("/api/v1/videos", response_model=VideoListResponse)
+    def list_videos() -> VideoListResponse:
+        return VideoListResponse(
+            videos=[_video_summary(row, service_settings) for row in video_repository.list_videos()]
+        )
+
+    @application.get("/api/v1/videos/{video_hash}", response_model=VideoDetailResponse)
+    def get_video(video_hash: str = SHA256_PATH) -> VideoDetailResponse:
+        row = video_repository.get_video(video_hash)
+        if row is None:
+            raise HTTPException(status_code=404, detail="Video was not found.")
+        return VideoDetailResponse(
+            video=_video_summary(row, service_settings),
+            boundaries=video_repository.get_document(video_hash),
+        )
+
+    @application.get("/api/v1/videos/{video_hash}/content")
+    def stream_video(video_hash: str = SHA256_PATH) -> FileResponse:
+        row = video_repository.get_video(video_hash)
+        if row is None:
+            raise HTTPException(status_code=404, detail="Video was not found.")
+        media_path = service_settings.videos_dir / row["stored_filename"]
+        if not media_path.is_file():
+            raise HTTPException(status_code=404, detail="Stored video file is missing.")
+        return FileResponse(
+            media_path,
+            media_type=row["content_type"],
+            headers={"Accept-Ranges": "bytes", "Cache-Control": "public, max-age=3600"},
+        )
 
     @application.get(
         "/api/v1/videos/{video_hash}/sentence-boundaries",
@@ -36,16 +166,8 @@ def create_app(store: BoundaryStore = None) -> FastAPI:
         response_model=SentenceBoundaryDocument,
         include_in_schema=False,
     )
-    def get_sentence_boundaries(
-        video_hash: str = ApiPath(
-            min_length=64,
-            max_length=64,
-            pattern=r"^[A-Fa-f0-9]{64}$",
-            description="SHA-256 hex digest of the exact video bytes",
-        ),
-        boundary_store: BoundaryStore = Depends(get_store),
-    ) -> SentenceBoundaryDocument:
-        document = boundary_store.get(video_hash)
+    def get_sentence_boundaries(video_hash: str = SHA256_PATH) -> SentenceBoundaryDocument:
+        document = find_document(video_hash)
         if document is None:
             raise HTTPException(
                 status_code=404,
@@ -56,7 +178,321 @@ def create_app(store: BoundaryStore = None) -> FastAPI:
             )
         return document
 
+    @application.post(
+        "/api/v1/admin/videos",
+        response_model=VideoUploadResponse,
+        status_code=202,
+        dependencies=[Depends(require_admin)],
+    )
+    async def upload_video(
+        background_tasks: BackgroundTasks,
+        file: UploadFile = File(...),
+        title: Optional[str] = Form(default=None),
+        language: Optional[str] = Form(default=None),
+    ) -> VideoUploadResponse:
+        original_name = Path(file.filename or "video.mp4").name
+        suffix = _validate_video_extension(original_name)
+        temporary_path = service_settings.work_dir / f"upload-{uuid.uuid4().hex}.part"
+        try:
+            video_hash, size_bytes = await run_in_threadpool(
+                _persist_upload,
+                file.file,
+                temporary_path,
+                service_settings.max_upload_bytes,
+            )
+        except UploadTooLargeError as exc:
+            temporary_path.unlink(missing_ok=True)
+            raise HTTPException(status_code=413, detail=str(exc)) from exc
+        except ValueError as exc:
+            temporary_path.unlink(missing_ok=True)
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+        finally:
+            await file.close()
+
+        content_type = file.content_type or mimetypes.guess_type(original_name)[0] or "video/mp4"
+        response = _register_uploaded_video(
+            settings=service_settings,
+            repository=video_repository,
+            temporary_path=temporary_path,
+            video_hash=video_hash,
+            size_bytes=size_bytes,
+            original_name=original_name,
+            suffix=suffix,
+            content_type=content_type,
+            title=title,
+            language=language,
+        )
+        background_tasks.add_task(_process_safely, processor, video_hash)
+        return response
+
+    @application.put(
+        "/api/v1/admin/videos/raw",
+        response_model=VideoUploadResponse,
+        status_code=202,
+        dependencies=[Depends(require_admin)],
+    )
+    async def upload_video_stream(
+        request: Request,
+        background_tasks: BackgroundTasks,
+        filename: str = Query(min_length=1, max_length=500),
+        title: Optional[str] = Query(default=None, max_length=200),
+        language: Optional[str] = Query(default=None, max_length=32),
+    ) -> VideoUploadResponse:
+        original_name = Path(filename).name
+        suffix = _validate_video_extension(original_name)
+        content_length = request.headers.get("content-length")
+        if content_length:
+            try:
+                declared_size = int(content_length)
+            except ValueError as exc:
+                raise HTTPException(status_code=400, detail="Invalid Content-Length header.") from exc
+            if declared_size > service_settings.max_upload_bytes:
+                raise HTTPException(status_code=413, detail="Upload exceeds the configured limit.")
+        temporary_path = service_settings.work_dir / f"upload-{uuid.uuid4().hex}.part"
+        try:
+            video_hash, size_bytes = await _persist_request_stream(
+                request,
+                temporary_path,
+                service_settings.max_upload_bytes,
+            )
+        except UploadTooLargeError as exc:
+            raise HTTPException(status_code=413, detail=str(exc)) from exc
+        except ValueError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+        response = _register_uploaded_video(
+            settings=service_settings,
+            repository=video_repository,
+            temporary_path=temporary_path,
+            video_hash=video_hash,
+            size_bytes=size_bytes,
+            original_name=original_name,
+            suffix=suffix,
+            content_type=request.headers.get("content-type") or "application/octet-stream",
+            title=title,
+            language=language,
+        )
+        background_tasks.add_task(_process_safely, processor, video_hash)
+        return response
+
+    @application.post(
+        "/api/v1/admin/videos/{video_hash}/process",
+        status_code=202,
+        dependencies=[Depends(require_admin)],
+    )
+    def reprocess_video(
+        background_tasks: BackgroundTasks,
+        video_hash: str = SHA256_PATH,
+    ) -> Dict[str, str]:
+        if video_repository.get_video(video_hash) is None:
+            raise HTTPException(status_code=404, detail="Video was not found.")
+        background_tasks.add_task(_process_safely, processor, video_hash)
+        return {"video_hash": video_hash.lower(), "status": "processing_queued"}
+
+    @application.put(
+        "/api/v1/admin/videos/{video_hash}/sentences/{sentence_index}",
+        response_model=SentenceBoundary,
+        dependencies=[Depends(require_admin)],
+    )
+    def update_sentence(
+        payload: SentenceTextUpdate,
+        video_hash: str = SHA256_PATH,
+        sentence_index: int = ApiPath(ge=0),
+    ) -> SentenceBoundary:
+        sentence = video_repository.update_sentence_text(
+            video_hash,
+            sentence_index,
+            payload.text,
+            payload.language,
+        )
+        if sentence is None:
+            raise HTTPException(status_code=404, detail="Sentence was not found.")
+        return sentence
+
+    @application.delete(
+        "/api/v1/admin/videos/{video_hash}",
+        status_code=204,
+        dependencies=[Depends(require_admin)],
+    )
+    def delete_video(video_hash: str = SHA256_PATH) -> Response:
+        row = video_repository.delete_video(video_hash)
+        if row is None:
+            raise HTTPException(status_code=404, detail="Video was not found.")
+        (service_settings.videos_dir / row["stored_filename"]).unlink(missing_ok=True)
+        return Response(status_code=204)
+
+    @application.post(
+        "/api/v1/videos/{video_hash}/sentences/{sentence_index}/assessments",
+        response_model=AssessmentResult,
+        dependencies=[Depends(require_client)],
+    )
+    async def assess_sentence(
+        video_hash: str = SHA256_PATH,
+        sentence_index: int = ApiPath(ge=0),
+        audio: UploadFile = File(...),
+        language: Optional[str] = Form(default=None),
+    ) -> AssessmentResult:
+        document = find_document(video_hash)
+        if document is None:
+            raise HTTPException(status_code=404, detail="Video or sentence boundaries were not found.")
+        if not moss.available:
+            raise HTTPException(status_code=503, detail="MOSS transcription is not configured.")
+
+        suffix = Path(audio.filename or "attempt.wav").suffix.lower() or ".wav"
+        temporary_name = f"attempt-{uuid.uuid4().hex}{suffix}"
+        temporary_path = service_settings.attempts_dir / temporary_name
+        completed = False
+        try:
+            await run_in_threadpool(
+                _persist_upload,
+                audio.file,
+                temporary_path,
+                service_settings.max_attempt_bytes,
+            )
+            result = await run_in_threadpool(
+                assessment_service.assess,
+                document=document,
+                sentence_index=sentence_index,
+                audio_path=temporary_path,
+                language=language,
+                retained_audio_filename=temporary_name if service_settings.keep_attempt_audio else None,
+            )
+            completed = True
+            return result
+        except UploadTooLargeError as exc:
+            raise HTTPException(status_code=413, detail=str(exc)) from exc
+        except LookupError as exc:
+            raise HTTPException(status_code=404, detail=str(exc)) from exc
+        except (ValueError, AudioAnalysisError) as exc:
+            raise HTTPException(status_code=422, detail=str(exc)) from exc
+        except RuntimeError as exc:
+            raise HTTPException(status_code=502, detail=str(exc)) from exc
+        finally:
+            await audio.close()
+            if not service_settings.keep_attempt_audio or not completed:
+                temporary_path.unlink(missing_ok=True)
+
     return application
 
 
+def _persist_upload(source: BinaryIO, destination: Path, max_bytes: int) -> tuple[str, int]:
+    digest = hashlib.sha256()
+    size = 0
+    destination.parent.mkdir(parents=True, exist_ok=True)
+    try:
+        with destination.open("wb") as output:
+            while True:
+                chunk = source.read(4 * 1024 * 1024)
+                if not chunk:
+                    break
+                size += len(chunk)
+                if size > max_bytes:
+                    raise UploadTooLargeError(f"Upload exceeds the {max_bytes}-byte limit.")
+                digest.update(chunk)
+                output.write(chunk)
+        if size == 0:
+            raise ValueError("Uploaded file is empty.")
+    except Exception:
+        destination.unlink(missing_ok=True)
+        raise
+    return digest.hexdigest(), size
+
+
+async def _persist_request_stream(
+    request: Request,
+    destination: Path,
+    max_bytes: int,
+) -> tuple[str, int]:
+    digest = hashlib.sha256()
+    size = 0
+    destination.parent.mkdir(parents=True, exist_ok=True)
+    try:
+        with destination.open("wb") as output:
+            async for chunk in request.stream():
+                if not chunk:
+                    continue
+                size += len(chunk)
+                if size > max_bytes:
+                    raise UploadTooLargeError(f"Upload exceeds the {max_bytes}-byte limit.")
+                digest.update(chunk)
+                output.write(chunk)
+        if size == 0:
+            raise ValueError("Uploaded file is empty.")
+    except Exception:
+        destination.unlink(missing_ok=True)
+        raise
+    return digest.hexdigest(), size
+
+
+def _validate_video_extension(filename: str) -> str:
+    suffix = Path(filename).suffix.lower()
+    if suffix not in ALLOWED_VIDEO_EXTENSIONS:
+        raise HTTPException(
+            status_code=415,
+            detail=f"Unsupported video extension: {suffix or '(none)'}",
+        )
+    return suffix
+
+
+def _register_uploaded_video(
+    *,
+    settings: Settings,
+    repository: VideoRepository,
+    temporary_path: Path,
+    video_hash: str,
+    size_bytes: int,
+    original_name: str,
+    suffix: str,
+    content_type: str,
+    title: Optional[str],
+    language: Optional[str],
+) -> VideoUploadResponse:
+    stored_filename = f"{video_hash}{suffix}"
+    destination = settings.videos_dir / stored_filename
+    if destination.exists():
+        temporary_path.unlink(missing_ok=True)
+    else:
+        temporary_path.replace(destination)
+    repository.upsert_upload(
+        video_hash=video_hash,
+        title=(title or Path(original_name).stem).strip() or video_hash,
+        filename=original_name,
+        stored_filename=stored_filename,
+        content_type=content_type,
+        size_bytes=size_bytes,
+        language=language.strip() if language else None,
+    )
+    return VideoUploadResponse(
+        video_hash=video_hash,
+        status="uploaded",
+        detail_url=f"/api/v1/videos/{video_hash}",
+    )
+
+
+def _process_safely(processor: VideoProcessor, video_hash: str) -> None:
+    try:
+        processor.process(video_hash)
+    except Exception:
+        logger.exception("Video processing failed for %s", video_hash)
+
+
+def _video_summary(row: Dict[str, Any], settings: Settings) -> VideoSummary:
+    relative_stream_url = f"/api/v1/videos/{row['video_hash']}/content"
+    stream_url = f"{settings.public_base_url}{relative_stream_url}" if settings.public_base_url else relative_stream_url
+    return VideoSummary(
+        video_hash=row["video_hash"],
+        title=row["title"],
+        filename=row["filename"],
+        content_type=row["content_type"],
+        size_bytes=row["size_bytes"],
+        duration_ms=row["duration_ms"],
+        language=row["language"],
+        status=row["status"],
+        error_message=row["error_message"],
+        sentence_count=row["sentence_count"],
+        stream_url=stream_url,
+        created_at=row["created_at"],
+        updated_at=row["updated_at"],
+    )
+
+
 app = create_app()
diff --git a/sentence_api/models.py b/sentence_api/models.py
index 095e738..0dd7bbc 100644
--- a/sentence_api/models.py
+++ b/sentence_api/models.py
@@ -1,4 +1,5 @@
-from typing import List, Optional
+from datetime import datetime
+from typing import Dict, List, Literal, Optional
 
 from pydantic import BaseModel, ConfigDict, Field, field_validator, model_validator
 
@@ -10,6 +11,8 @@ class SentenceBoundary(BaseModel):
     start_ms: int = Field(ge=0)
     end_ms: int = Field(gt=0)
     text: Optional[str] = None
+    language: Optional[str] = None
+    reference_speech_duration_ms: Optional[int] = Field(default=None, gt=0)
 
     @model_validator(mode="after")
     def validate_range(self):
@@ -46,3 +49,79 @@ class SentenceBoundaryDocument(BaseModel):
                 raise ValueError("sentence end_ms cannot exceed duration_ms")
             previous_end = sentence.end_ms
         return self
+
+
+VideoStatus = Literal["uploaded", "processing", "ready", "failed"]
+
+
+class VideoSummary(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    video_hash: str
+    title: str
+    filename: str
+    content_type: str
+    size_bytes: int = Field(ge=0)
+    duration_ms: Optional[int] = Field(default=None, gt=0)
+    language: Optional[str] = None
+    status: VideoStatus
+    error_message: Optional[str] = None
+    sentence_count: int = Field(ge=0)
+    stream_url: str
+    created_at: datetime
+    updated_at: datetime
+
+
+class VideoListResponse(BaseModel):
+    videos: List[VideoSummary]
+
+
+class VideoUploadResponse(BaseModel):
+    video_hash: str
+    status: VideoStatus
+    detail_url: str
+
+
+class VideoDetailResponse(BaseModel):
+    video: VideoSummary
+    boundaries: Optional[SentenceBoundaryDocument] = None
+
+
+class SentenceTextUpdate(BaseModel):
+    text: str = Field(min_length=1, max_length=4000)
+    language: Optional[str] = Field(default=None, max_length=32)
+
+
+class TextSubstitution(BaseModel):
+    expected: str
+    actual: str
+
+
+class AssessmentResult(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    attempt_id: str
+    scoring_version: str
+    overall_score: float = Field(ge=0, le=100)
+    passed: bool
+    pass_score: float = Field(ge=0, le=100)
+    content_score: float = Field(ge=0, le=100)
+    completeness_score: float = Field(ge=0, le=100)
+    fluency_score: float = Field(ge=0, le=100)
+    pronunciation_score: Optional[float] = Field(default=None, ge=0, le=100)
+    prosody_score: Optional[float] = Field(default=None, ge=0, le=100)
+    duration_score: float = Field(ge=0, le=100)
+    pause_score: float = Field(ge=0, le=100)
+    speech_rate_score: float = Field(ge=0, le=100)
+    reference_text: str
+    recognized_text: str
+    reference_duration_ms: int = Field(gt=0)
+    reference_speech_duration_ms: int = Field(gt=0)
+    student_recording_duration_ms: int = Field(gt=0)
+    student_speech_duration_ms: int = Field(gt=0)
+    duration_ratio: float = Field(gt=0)
+    missing_tokens: List[str]
+    extra_tokens: List[str]
+    substitutions: List[TextSubstitution]
+    feedback: str
+    details: Dict[str, str] = Field(default_factory=dict)
diff --git a/sentence_api/nginx.conf.example b/sentence_api/nginx.conf.example
new file mode 100644
index 0000000..8f661ea
--- /dev/null
+++ b/sentence_api/nginx.conf.example
@@ -0,0 +1,30 @@
+server {
+    listen 80;
+    server_name video_service.d1kt.cn;
+
+    client_max_body_size 12g;
+    client_body_timeout 7200s;
+    send_timeout 7200s;
+
+    location / {
+        proxy_pass http://127.0.0.1:8000;
+        proxy_http_version 1.1;
+        proxy_set_header Host $host;
+        proxy_set_header X-Real-IP $remote_addr;
+        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+        proxy_set_header X-Forwarded-Proto $scheme;
+        proxy_request_buffering off;
+        proxy_read_timeout 7200s;
+        proxy_send_timeout 7200s;
+    }
+
+    location ~ ^/api/v1/videos/[0-9a-fA-F]{64}/content$ {
+        proxy_pass http://127.0.0.1:8000;
+        proxy_http_version 1.1;
+        proxy_set_header Host $host;
+        proxy_set_header Range $http_range;
+        proxy_set_header If-Range $http_if_range;
+        proxy_force_ranges on;
+        proxy_buffering off;
+    }
+}
diff --git a/sentence_api/processing.py b/sentence_api/processing.py
new file mode 100644
index 0000000..2bfee26
--- /dev/null
+++ b/sentence_api/processing.py
@@ -0,0 +1,144 @@
+import shutil
+import subprocess
+import uuid
+from pathlib import Path
+from typing import List, Optional
+
+from .audio_metrics import AudioAnalysisError, analyze_samples, decode_audio_mono
+from .config import Settings
+from .generate_boundaries import ALGORITHM_VERSION, make_entry
+from .models import SentenceBoundary, SentenceBoundaryDocument
+from .repository import VideoRepository
+from .transcription import Transcript, Transcriber
+
+
+MOSS_ALGORITHM_VERSION = "moss-timestamp-v1"
+
+
+class VideoProcessor:
+    def __init__(
+        self,
+        settings: Settings,
+        repository: VideoRepository,
+        transcriber: Transcriber,
+    ):
+        self.settings = settings
+        self.repository = repository
+        self.transcriber = transcriber
+
+    def process(self, video_hash: str) -> None:
+        video = self.repository.get_video(video_hash)
+        if video is None:
+            raise ValueError(f"Unknown video: {video_hash}")
+        media_path = self.settings.videos_dir / video["stored_filename"]
+        if not media_path.is_file():
+            raise FileNotFoundError(f"Stored video is missing: {media_path.name}")
+
+        self.repository.mark_processing(video_hash)
+        work_path: Optional[Path] = None
+        try:
+            if self.transcriber.available:
+                work_path = self.settings.work_dir / f"{video_hash}-{uuid.uuid4().hex}.wav"
+                extract_audio(media_path, work_path)
+                transcript = self.transcriber.transcribe(work_path, video.get("language"))
+                document = document_from_transcript(
+                    video_hash=video_hash,
+                    duration_ms=_media_duration_ms(media_path),
+                    transcript=transcript,
+                    language=video.get("language"),
+                    audio_path=work_path,
+                )
+                if not document.sentences:
+                    raise RuntimeError("MOSS returned no timestamped speech segments.")
+                self.repository.save_processing_result(document, transcript.text)
+            else:
+                entry, _ = make_entry(media_path, video_hash=video_hash)
+                document = SentenceBoundaryDocument(
+                    video_hash=video_hash,
+                    duration_ms=entry["duration_ms"],
+                    algorithm_version=ALGORITHM_VERSION,
+                    sentences=entry["sentences"],
+                )
+                self.repository.save_processing_result(document, None)
+        except Exception as exc:
+            self.repository.mark_failed(video_hash, str(exc))
+            raise
+        finally:
+            if work_path is not None:
+                work_path.unlink(missing_ok=True)
+
+
+def extract_audio(media_path: Path, output_path: Path) -> None:
+    ffmpeg = shutil.which("ffmpeg")
+    if ffmpeg is None:
+        raise RuntimeError("ffmpeg is required for MOSS transcription but was not found.")
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    command = [
+        ffmpeg,
+        "-hide_banner",
+        "-loglevel",
+        "error",
+        "-y",
+        "-i",
+        str(media_path),
+        "-vn",
+        "-ac",
+        "1",
+        "-ar",
+        "16000",
+        "-c:a",
+        "pcm_s16le",
+        str(output_path),
+    ]
+    completed = subprocess.run(command, capture_output=True, text=True, timeout=7200)
+    if completed.returncode != 0:
+        message = completed.stderr.strip() or "unknown ffmpeg error"
+        raise RuntimeError(f"Could not extract video audio: {message[-2000:]}")
+
+
+def document_from_transcript(
+    *,
+    video_hash: str,
+    duration_ms: int,
+    transcript: Transcript,
+    language: Optional[str],
+    audio_path: Path,
+) -> SentenceBoundaryDocument:
+    samples, sample_rate = decode_audio_mono(audio_path)
+    sentences: List[SentenceBoundary] = []
+    previous_end = 0
+    for segment in sorted(transcript.segments, key=lambda item: (item.start_seconds, item.end_seconds)):
+        start_ms = max(previous_end, int(round(segment.start_seconds * 1000)))
+        end_ms = min(duration_ms, int(round(segment.end_seconds * 1000)))
+        if not segment.text.strip() or end_ms <= start_ms:
+            continue
+        start_sample = max(0, int(start_ms / 1000 * sample_rate))
+        end_sample = min(samples.size, int(end_ms / 1000 * sample_rate))
+        try:
+            metrics = analyze_samples(samples[start_sample:end_sample], sample_rate)
+            speech_duration_ms = metrics.speech_duration_ms
+        except AudioAnalysisError:
+            speech_duration_ms = end_ms - start_ms
+        sentences.append(
+            SentenceBoundary(
+                index=len(sentences),
+                start_ms=start_ms,
+                end_ms=end_ms,
+                text=segment.text.strip(),
+                language=language,
+                reference_speech_duration_ms=max(1, speech_duration_ms),
+            )
+        )
+        previous_end = end_ms
+    return SentenceBoundaryDocument(
+        video_hash=video_hash,
+        duration_ms=duration_ms,
+        algorithm_version=MOSS_ALGORITHM_VERSION,
+        sentences=sentences,
+    )
+
+
+def _media_duration_ms(path: Path) -> int:
+    from .generate_boundaries import media_duration_ms
+
+    return media_duration_ms(path)
diff --git a/sentence_api/repository.py b/sentence_api/repository.py
new file mode 100644
index 0000000..c41b4e6
--- /dev/null
+++ b/sentence_api/repository.py
@@ -0,0 +1,305 @@
+import json
+import sqlite3
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+from .models import SentenceBoundary, SentenceBoundaryDocument
+from .store import normalize_video_hash
+
+
+def utc_now() -> str:
+    return datetime.now(timezone.utc).isoformat()
+
+
+class VideoRepository:
+    def __init__(self, database_path: Path):
+        self.database_path = Path(database_path)
+        self.database_path.parent.mkdir(parents=True, exist_ok=True)
+        self._initialize()
+
+    def _connect(self) -> sqlite3.Connection:
+        connection = sqlite3.connect(self.database_path, timeout=30)
+        connection.row_factory = sqlite3.Row
+        connection.execute("PRAGMA foreign_keys = ON")
+        connection.execute("PRAGMA journal_mode = WAL")
+        return connection
+
+    def _initialize(self) -> None:
+        with self._connect() as connection:
+            connection.executescript(
+                """
+                CREATE TABLE IF NOT EXISTS videos (
+                    video_hash TEXT PRIMARY KEY,
+                    title TEXT NOT NULL,
+                    filename TEXT NOT NULL,
+                    stored_filename TEXT NOT NULL,
+                    content_type TEXT NOT NULL,
+                    size_bytes INTEGER NOT NULL,
+                    duration_ms INTEGER,
+                    language TEXT,
+                    status TEXT NOT NULL,
+                    error_message TEXT,
+                    algorithm_version TEXT,
+                    transcription TEXT,
+                    created_at TEXT NOT NULL,
+                    updated_at TEXT NOT NULL
+                );
+
+                CREATE TABLE IF NOT EXISTS sentences (
+                    video_hash TEXT NOT NULL,
+                    sentence_index INTEGER NOT NULL,
+                    start_ms INTEGER NOT NULL,
+                    end_ms INTEGER NOT NULL,
+                    text TEXT,
+                    language TEXT,
+                    reference_speech_duration_ms INTEGER,
+                    PRIMARY KEY (video_hash, sentence_index),
+                    FOREIGN KEY (video_hash) REFERENCES videos(video_hash) ON DELETE CASCADE
+                );
+
+                CREATE TABLE IF NOT EXISTS attempts (
+                    attempt_id TEXT PRIMARY KEY,
+                    video_hash TEXT NOT NULL,
+                    sentence_index INTEGER NOT NULL,
+                    result_json TEXT NOT NULL,
+                    audio_filename TEXT,
+                    created_at TEXT NOT NULL,
+                    FOREIGN KEY (video_hash) REFERENCES videos(video_hash) ON DELETE CASCADE
+                );
+                """
+            )
+
+    def upsert_upload(
+        self,
+        *,
+        video_hash: str,
+        title: str,
+        filename: str,
+        stored_filename: str,
+        content_type: str,
+        size_bytes: int,
+        language: Optional[str],
+    ) -> None:
+        normalized_hash = normalize_video_hash(video_hash)
+        now = utc_now()
+        with self._connect() as connection:
+            existing = connection.execute(
+                "SELECT created_at FROM videos WHERE video_hash = ?",
+                (normalized_hash,),
+            ).fetchone()
+            created_at = existing["created_at"] if existing else now
+            connection.execute(
+                """
+                INSERT INTO videos (
+                    video_hash, title, filename, stored_filename, content_type,
+                    size_bytes, duration_ms, language, status, error_message,
+                    algorithm_version, transcription, created_at, updated_at
+                ) VALUES (?, ?, ?, ?, ?, ?, NULL, ?, 'uploaded', NULL, NULL, NULL, ?, ?)
+                ON CONFLICT(video_hash) DO UPDATE SET
+                    title = excluded.title,
+                    filename = excluded.filename,
+                    stored_filename = excluded.stored_filename,
+                    content_type = excluded.content_type,
+                    size_bytes = excluded.size_bytes,
+                    language = excluded.language,
+                    status = 'uploaded',
+                    error_message = NULL,
+                    updated_at = excluded.updated_at
+                """,
+                (
+                    normalized_hash,
+                    title,
+                    filename,
+                    stored_filename,
+                    content_type,
+                    size_bytes,
+                    language,
+                    created_at,
+                    now,
+                ),
+            )
+
+    def mark_processing(self, video_hash: str) -> None:
+        self._update_status(video_hash, "processing", None)
+
+    def mark_failed(self, video_hash: str, message: str) -> None:
+        self._update_status(video_hash, "failed", message[:4000])
+
+    def _update_status(self, video_hash: str, status: str, error_message: Optional[str]) -> None:
+        with self._connect() as connection:
+            connection.execute(
+                "UPDATE videos SET status = ?, error_message = ?, updated_at = ? WHERE video_hash = ?",
+                (status, error_message, utc_now(), normalize_video_hash(video_hash)),
+            )
+
+    def save_processing_result(
+        self,
+        document: SentenceBoundaryDocument,
+        transcription: Optional[str],
+    ) -> None:
+        with self._connect() as connection:
+            connection.execute(
+                "DELETE FROM sentences WHERE video_hash = ?",
+                (document.video_hash,),
+            )
+            connection.executemany(
+                """
+                INSERT INTO sentences (
+                    video_hash, sentence_index, start_ms, end_ms, text,
+                    language, reference_speech_duration_ms
+                ) VALUES (?, ?, ?, ?, ?, ?, ?)
+                """,
+                [
+                    (
+                        document.video_hash,
+                        sentence.index,
+                        sentence.start_ms,
+                        sentence.end_ms,
+                        sentence.text,
+                        sentence.language,
+                        sentence.reference_speech_duration_ms,
+                    )
+                    for sentence in document.sentences
+                ],
+            )
+            connection.execute(
+                """
+                UPDATE videos SET duration_ms = ?, status = 'ready', error_message = NULL,
+                    algorithm_version = ?, transcription = ?, updated_at = ?
+                WHERE video_hash = ?
+                """,
+                (
+                    document.duration_ms,
+                    document.algorithm_version,
+                    transcription,
+                    utc_now(),
+                    document.video_hash,
+                ),
+            )
+
+    def list_videos(self) -> List[Dict[str, Any]]:
+        with self._connect() as connection:
+            rows = connection.execute(
+                """
+                SELECT v.*, COUNT(s.sentence_index) AS sentence_count
+                FROM videos v
+                LEFT JOIN sentences s ON s.video_hash = v.video_hash
+                GROUP BY v.video_hash
+                ORDER BY v.created_at DESC
+                """
+            ).fetchall()
+        return [dict(row) for row in rows]
+
+    def get_video(self, video_hash: str) -> Optional[Dict[str, Any]]:
+        with self._connect() as connection:
+            row = connection.execute(
+                """
+                SELECT v.*, COUNT(s.sentence_index) AS sentence_count
+                FROM videos v
+                LEFT JOIN sentences s ON s.video_hash = v.video_hash
+                WHERE v.video_hash = ?
+                GROUP BY v.video_hash
+                """,
+                (normalize_video_hash(video_hash),),
+            ).fetchone()
+        return dict(row) if row else None
+
+    def get_document(self, video_hash: str) -> Optional[SentenceBoundaryDocument]:
+        video = self.get_video(video_hash)
+        if video is None or video["duration_ms"] is None:
+            return None
+        with self._connect() as connection:
+            rows = connection.execute(
+                "SELECT * FROM sentences WHERE video_hash = ? ORDER BY sentence_index",
+                (normalize_video_hash(video_hash),),
+            ).fetchall()
+        return SentenceBoundaryDocument(
+            video_hash=video["video_hash"],
+            duration_ms=video["duration_ms"],
+            algorithm_version=video["algorithm_version"] or "unknown",
+            sentences=[
+                SentenceBoundary(
+                    index=row["sentence_index"],
+                    start_ms=row["start_ms"],
+                    end_ms=row["end_ms"],
+                    text=row["text"],
+                    language=row["language"],
+                    reference_speech_duration_ms=row["reference_speech_duration_ms"],
+                )
+                for row in rows
+            ],
+        )
+
+    def update_sentence_text(
+        self,
+        video_hash: str,
+        sentence_index: int,
+        text: str,
+        language: Optional[str],
+    ) -> Optional[SentenceBoundary]:
+        normalized_hash = normalize_video_hash(video_hash)
+        with self._connect() as connection:
+            cursor = connection.execute(
+                """
+                UPDATE sentences SET text = ?, language = COALESCE(?, language)
+                WHERE video_hash = ? AND sentence_index = ?
+                """,
+                (text.strip(), language, normalized_hash, sentence_index),
+            )
+            if cursor.rowcount == 0:
+                return None
+            connection.execute(
+                "UPDATE videos SET updated_at = ? WHERE video_hash = ?",
+                (utc_now(), normalized_hash),
+            )
+            row = connection.execute(
+                "SELECT * FROM sentences WHERE video_hash = ? AND sentence_index = ?",
+                (normalized_hash, sentence_index),
+            ).fetchone()
+        return SentenceBoundary(
+            index=row["sentence_index"],
+            start_ms=row["start_ms"],
+            end_ms=row["end_ms"],
+            text=row["text"],
+            language=row["language"],
+            reference_speech_duration_ms=row["reference_speech_duration_ms"],
+        )
+
+    def record_attempt(
+        self,
+        *,
+        attempt_id: str,
+        video_hash: str,
+        sentence_index: int,
+        result: Dict[str, Any],
+        audio_filename: Optional[str],
+    ) -> None:
+        with self._connect() as connection:
+            connection.execute(
+                """
+                INSERT INTO attempts (
+                    attempt_id, video_hash, sentence_index, result_json,
+                    audio_filename, created_at
+                ) VALUES (?, ?, ?, ?, ?, ?)
+                """,
+                (
+                    attempt_id,
+                    normalize_video_hash(video_hash),
+                    sentence_index,
+                    json.dumps(result, ensure_ascii=False),
+                    audio_filename,
+                    utc_now(),
+                ),
+            )
+
+    def delete_video(self, video_hash: str) -> Optional[Dict[str, Any]]:
+        video = self.get_video(video_hash)
+        if video is None:
+            return None
+        with self._connect() as connection:
+            connection.execute(
+                "DELETE FROM videos WHERE video_hash = ?",
+                (normalize_video_hash(video_hash),),
+            )
+        return video
diff --git a/sentence_api/requirements.txt b/sentence_api/requirements.txt
index 6f55850..a310817 100644
--- a/sentence_api/requirements.txt
+++ b/sentence_api/requirements.txt
@@ -3,5 +3,6 @@ uvicorn[standard]>=0.30,<1
 pydantic>=2.7,<3
 pytest>=8,<9
 httpx>=0.27,<1
+python-multipart>=0.0.9,<1
 av>=12.0
 numpy>=1.26
diff --git a/sentence_api/scoring.py b/sentence_api/scoring.py
new file mode 100644
index 0000000..f9a7423
--- /dev/null
+++ b/sentence_api/scoring.py
@@ -0,0 +1,241 @@
+import math
+import unicodedata
+from dataclasses import dataclass
+from typing import List, Sequence, Tuple
+
+from .audio_metrics import AudioMetrics
+
+
+SCORING_VERSION = "asr-fluency-v1"
+
+
+@dataclass(frozen=True)
+class TextAlignment:
+    reference_tokens: List[str]
+    recognized_tokens: List[str]
+    matches: int
+    missing_tokens: List[str]
+    extra_tokens: List[str]
+    substitutions: List[Tuple[str, str]]
+    content_score: float
+    completeness_score: float
+
+
+@dataclass(frozen=True)
+class ScoreBreakdown:
+    overall_score: float
+    content_score: float
+    completeness_score: float
+    fluency_score: float
+    duration_score: float
+    pause_score: float
+    speech_rate_score: float
+    duration_ratio: float
+    missing_tokens: List[str]
+    extra_tokens: List[str]
+    substitutions: List[Tuple[str, str]]
+    feedback: str
+
+
+def tokenize(text: str) -> List[str]:
+    normalized = unicodedata.normalize("NFKC", text).lower().strip()
+    tokens: List[str] = []
+    word: List[str] = []
+
+    def flush_word() -> None:
+        if word:
+            token = "".join(word).strip("'")
+            if token:
+                tokens.append(token)
+            word.clear()
+
+    for char in normalized:
+        if _is_cjk(char):
+            flush_word()
+            tokens.append(char)
+        elif char.isalnum() or (char == "'" and word):
+            word.append(char)
+        else:
+            flush_word()
+    flush_word()
+    return tokens
+
+
+def align_text(reference_text: str, recognized_text: str) -> TextAlignment:
+    reference = tokenize(reference_text)
+    recognized = tokenize(recognized_text)
+    if not reference:
+        raise ValueError("Reference text does not contain any scoreable tokens.")
+
+    costs = [[0] * (len(recognized) + 1) for _ in range(len(reference) + 1)]
+    for row in range(1, len(reference) + 1):
+        costs[row][0] = row
+    for column in range(1, len(recognized) + 1):
+        costs[0][column] = column
+    for row in range(1, len(reference) + 1):
+        for column in range(1, len(recognized) + 1):
+            substitution_cost = 0 if reference[row - 1] == recognized[column - 1] else 1
+            costs[row][column] = min(
+                costs[row - 1][column] + 1,
+                costs[row][column - 1] + 1,
+                costs[row - 1][column - 1] + substitution_cost,
+            )
+
+    row = len(reference)
+    column = len(recognized)
+    matches = 0
+    missing: List[str] = []
+    extra: List[str] = []
+    substitutions: List[Tuple[str, str]] = []
+    while row > 0 or column > 0:
+        if row > 0 and column > 0:
+            same = reference[row - 1] == recognized[column - 1]
+            diagonal_cost = costs[row - 1][column - 1] + (0 if same else 1)
+            if costs[row][column] == diagonal_cost:
+                if same:
+                    matches += 1
+                else:
+                    substitutions.append((reference[row - 1], recognized[column - 1]))
+                row -= 1
+                column -= 1
+                continue
+        if row > 0 and costs[row][column] == costs[row - 1][column] + 1:
+            missing.append(reference[row - 1])
+            row -= 1
+        else:
+            extra.append(recognized[column - 1])
+            column -= 1
+
+    missing.reverse()
+    extra.reverse()
+    substitutions.reverse()
+    distance = len(missing) + len(extra) + len(substitutions)
+    accuracy = max(0.0, 1.0 - distance / len(reference)) * 100
+    attempted_reference_tokens = matches + len(substitutions)
+    completeness = attempted_reference_tokens / len(reference) * 100
+    content = accuracy * 0.7 + completeness * 0.3
+    return TextAlignment(
+        reference_tokens=reference,
+        recognized_tokens=recognized,
+        matches=matches,
+        missing_tokens=missing,
+        extra_tokens=extra,
+        substitutions=substitutions,
+        content_score=_round_score(content),
+        completeness_score=_round_score(completeness),
+    )
+
+
+def score_attempt(
+    *,
+    reference_text: str,
+    recognized_text: str,
+    reference_speech_duration_ms: int,
+    student_metrics: AudioMetrics,
+) -> ScoreBreakdown:
+    if reference_speech_duration_ms <= 0:
+        raise ValueError("reference_speech_duration_ms must be positive")
+    alignment = align_text(reference_text, recognized_text)
+    ratio = student_metrics.speech_duration_ms / reference_speech_duration_ms
+    duration = duration_similarity_score(ratio)
+    pause = pause_score(student_metrics.internal_pause_ratio)
+
+    reference_rate = len(alignment.reference_tokens) / (reference_speech_duration_ms / 1000)
+    recognized_units = max(1, len(alignment.recognized_tokens))
+    student_rate = recognized_units / (student_metrics.speech_duration_ms / 1000)
+    rate_ratio = student_rate / reference_rate if reference_rate > 0 else 1.0
+    rate = symmetric_rate_score(rate_ratio)
+
+    fluency = duration * 0.35 + pause * 0.40 + rate * 0.25
+    overall = alignment.content_score * 0.80 + fluency * 0.20
+    feedback = build_feedback(alignment, ratio, student_metrics.internal_pause_ratio)
+    return ScoreBreakdown(
+        overall_score=_round_score(overall),
+        content_score=alignment.content_score,
+        completeness_score=alignment.completeness_score,
+        fluency_score=_round_score(fluency),
+        duration_score=_round_score(duration),
+        pause_score=_round_score(pause),
+        speech_rate_score=_round_score(rate),
+        duration_ratio=round(ratio, 4),
+        missing_tokens=alignment.missing_tokens,
+        extra_tokens=alignment.extra_tokens,
+        substitutions=alignment.substitutions,
+        feedback=feedback,
+    )
+
+
+def duration_similarity_score(ratio: float) -> float:
+    if ratio <= 0:
+        return 0.0
+    if 0.80 <= ratio <= 1.30:
+        return 100.0
+    if 0.65 <= ratio < 0.80:
+        return _interpolate(ratio, 0.65, 0.80, 60, 100)
+    if 1.30 < ratio <= 1.50:
+        return _interpolate(ratio, 1.30, 1.50, 100, 60)
+    if 0.50 <= ratio < 0.65:
+        return _interpolate(ratio, 0.50, 0.65, 20, 60)
+    if 1.50 < ratio <= 1.80:
+        return _interpolate(ratio, 1.50, 1.80, 60, 20)
+    if ratio < 0.50:
+        return max(0.0, ratio / 0.50 * 20)
+    return max(0.0, 20 - (ratio - 1.80) / 0.40 * 20)
+
+
+def pause_score(internal_pause_ratio: float) -> float:
+    if internal_pause_ratio <= 0.15:
+        return 100.0
+    if internal_pause_ratio <= 0.30:
+        return _interpolate(internal_pause_ratio, 0.15, 0.30, 100, 70)
+    if internal_pause_ratio <= 0.50:
+        return _interpolate(internal_pause_ratio, 0.30, 0.50, 70, 20)
+    return max(0.0, 20 - (internal_pause_ratio - 0.50) / 0.30 * 20)
+
+
+def symmetric_rate_score(ratio: float) -> float:
+    if ratio <= 0:
+        return 0.0
+    deviation = abs(math.log(ratio))
+    free_tolerance = math.log(1.20)
+    if deviation <= free_tolerance:
+        return 100.0
+    return max(0.0, 100 - (deviation - free_tolerance) / math.log(2.5) * 100)
+
+
+def build_feedback(
+    alignment: TextAlignment,
+    duration_ratio: float,
+    internal_pause_ratio: float,
+) -> str:
+    messages = []
+    if alignment.missing_tokens:
+        messages.append("存在漏读")
+    if alignment.substitutions:
+        messages.append("存在错读")
+    if alignment.extra_tokens:
+        messages.append("存在多读")
+    if duration_ratio < 0.65:
+        messages.append("朗读明显偏快或内容不完整")
+    elif duration_ratio > 1.50:
+        messages.append("朗读速度偏慢")
+    if internal_pause_ratio > 0.30:
+        messages.append("句内停顿偏多")
+    return "；".join(messages) if messages else "内容和朗读节奏匹配良好"
+
+
+def _is_cjk(char: str) -> bool:
+    codepoint = ord(char)
+    return (
+        0x3400 <= codepoint <= 0x4DBF
+        or 0x4E00 <= codepoint <= 0x9FFF
+        or 0xF900 <= codepoint <= 0xFAFF
+    )
+
+
+def _interpolate(value: float, start: float, end: float, start_score: float, end_score: float) -> float:
+    return start_score + (value - start) / (end - start) * (end_score - start_score)
+
+
+def _round_score(value: float) -> float:
+    return round(min(100.0, max(0.0, value)), 1)
diff --git a/sentence_api/static/admin.css b/sentence_api/static/admin.css
new file mode 100644
index 0000000..648e737
--- /dev/null
+++ b/sentence_api/static/admin.css
@@ -0,0 +1,117 @@
+:root {
+  color-scheme: light;
+  font-family: Inter, "PingFang SC", "Microsoft YaHei", sans-serif;
+  color: #17202a;
+  background: #f4f6f7;
+  letter-spacing: 0;
+}
+
+* { box-sizing: border-box; }
+body { margin: 0; min-width: 320px; }
+button, input, select { font: inherit; letter-spacing: 0; }
+button { cursor: pointer; }
+
+.topbar {
+  min-height: 72px;
+  padding: 14px 28px;
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 24px;
+  color: #fff;
+  background: #20262c;
+  border-bottom: 3px solid #1f8a70;
+}
+
+h1, h2 { margin: 0; letter-spacing: 0; }
+h1 { font-size: 21px; line-height: 1.35; }
+h2 { font-size: 17px; }
+.service-state, .section-heading span { font-size: 13px; color: #66727d; }
+.topbar .service-state { color: #c8d0d7; }
+.admin-auth { display: flex; align-items: center; gap: 8px; }
+.admin-auth label { font-size: 13px; color: #d5dadd; }
+
+input, select {
+  min-height: 36px;
+  padding: 7px 10px;
+  border: 1px solid #b8c0c7;
+  border-radius: 4px;
+  background: #fff;
+}
+
+button {
+  min-height: 36px;
+  padding: 7px 13px;
+  border: 1px solid #aab4bc;
+  border-radius: 4px;
+  color: #26323c;
+  background: #fff;
+}
+button:hover { background: #edf1f3; }
+button:disabled { cursor: default; opacity: .55; }
+.primary { color: #fff; border-color: #176b57; background: #1f8a70; }
+.primary:hover { background: #176b57; }
+.danger { color: #a12622; border-color: #d4aaa8; }
+.icon-button { width: 36px; padding: 0; font-weight: 700; }
+
+main { width: min(1480px, 100%); margin: 0 auto; }
+section { padding: 24px 28px; }
+.upload-band { background: #fff; border-bottom: 1px solid #dce1e5; }
+.video-band { min-height: 400px; }
+.section-heading {
+  min-height: 30px;
+  display: flex;
+  align-items: baseline;
+  justify-content: space-between;
+  gap: 16px;
+  margin-bottom: 14px;
+}
+
+#upload-form {
+  display: grid;
+  grid-template-columns: minmax(260px, 2fr) minmax(180px, 1fr) 150px auto;
+  align-items: end;
+  gap: 12px;
+}
+#upload-form label { display: grid; gap: 5px; font-size: 13px; color: #4f5b64; }
+#upload-progress { width: 100%; height: 8px; margin-top: 14px; accent-color: #1f8a70; }
+
+.table-scroll { overflow: auto; border: 1px solid #d7dde1; background: #fff; }
+table { width: 100%; border-collapse: collapse; font-size: 13px; }
+th, td { padding: 10px 12px; text-align: left; border-bottom: 1px solid #e3e7ea; vertical-align: middle; }
+th { position: sticky; top: 0; z-index: 1; color: #44515b; background: #edf1f3; font-weight: 600; }
+tr:last-child td { border-bottom: 0; }
+.actions-column { min-width: 280px; }
+.actions { display: flex; gap: 6px; white-space: nowrap; }
+.actions button, .actions a { min-height: 30px; padding: 5px 9px; font-size: 12px; }
+.actions a { display: inline-flex; align-items: center; border: 1px solid #aab4bc; border-radius: 4px; color: #26323c; text-decoration: none; }
+.status { display: inline-block; min-width: 64px; font-weight: 600; }
+.status-ready { color: #167356; }
+.status-failed { color: #b12d28; }
+.status-processing, .status-uploaded { color: #a45d08; }
+.error-message { max-width: 320px; margin-top: 3px; color: #a12622; font-size: 12px; white-space: normal; }
+.empty-state { padding: 70px 20px; text-align: center; color: #6f7a82; }
+
+dialog {
+  width: min(1200px, calc(100vw - 32px));
+  max-height: calc(100vh - 32px);
+  padding: 0;
+  border: 1px solid #909ba3;
+  border-radius: 6px;
+  box-shadow: 0 18px 60px rgb(0 0 0 / 25%);
+}
+dialog::backdrop { background: rgb(20 25 28 / 55%); }
+.dialog-heading { padding: 16px 18px; display: flex; justify-content: space-between; align-items: center; border-bottom: 1px solid #dce1e5; }
+.dialog-heading span { font-size: 12px; color: #68747d; }
+.sentence-scroll { max-height: calc(100vh - 120px); overflow: auto; }
+.sentence-scroll textarea { width: min(600px, 45vw); min-width: 260px; min-height: 58px; resize: vertical; padding: 7px; border: 1px solid #b8c0c7; border-radius: 3px; }
+.sentence-scroll select { min-width: 90px; }
+
+@media (max-width: 820px) {
+  .topbar { align-items: stretch; flex-direction: column; padding: 14px 16px; }
+  .admin-auth { flex-wrap: wrap; }
+  .admin-auth input { flex: 1 1 150px; }
+  section { padding: 20px 16px; }
+  #upload-form { grid-template-columns: 1fr; }
+  .actions-column { min-width: 250px; }
+}
diff --git a/sentence_api/static/admin.html b/sentence_api/static/admin.html
new file mode 100644
index 0000000..a5faf0b
--- /dev/null
+++ b/sentence_api/static/admin.html
@@ -0,0 +1,96 @@
+<!doctype html>
+<html lang="zh-CN">
+<head>
+  <meta charset="utf-8">
+  <meta name="viewport" content="width=device-width, initial-scale=1">
+  <title>口语训练视频管理</title>
+  <link rel="stylesheet" href="/static/admin.css">
+</head>
+<body>
+  <header class="topbar">
+    <div>
+      <h1>口语训练视频管理</h1>
+      <span id="service-state" class="service-state">检查服务</span>
+    </div>
+    <div class="admin-auth">
+      <label for="admin-key">管理密钥</label>
+      <input id="admin-key" type="password" autocomplete="current-password">
+      <button id="save-key" type="button">保存</button>
+      <button id="refresh" type="button">刷新</button>
+    </div>
+  </header>
+
+  <main>
+    <section class="upload-band" aria-labelledby="upload-heading">
+      <div class="section-heading">
+        <h2 id="upload-heading">上传视频</h2>
+        <span id="upload-state"></span>
+      </div>
+      <form id="upload-form">
+        <label class="file-field">
+          <span>视频文件</span>
+          <input id="video-file" name="file" type="file" accept="video/*" required>
+        </label>
+        <label>
+          <span>标题</span>
+          <input id="video-title" name="title" type="text" maxlength="200">
+        </label>
+        <label>
+          <span>语言</span>
+          <select id="video-language" name="language">
+            <option value="">自动识别</option>
+            <option value="en">英语</option>
+            <option value="zh">中文</option>
+          </select>
+        </label>
+        <button class="primary" type="submit">上传并处理</button>
+      </form>
+      <progress id="upload-progress" value="0" max="100" hidden></progress>
+    </section>
+
+    <section class="video-band" aria-labelledby="videos-heading">
+      <div class="section-heading">
+        <h2 id="videos-heading">视频库</h2>
+        <span id="video-count">0 个视频</span>
+      </div>
+      <div class="table-scroll">
+        <table>
+          <thead>
+            <tr>
+              <th>标题</th>
+              <th>状态</th>
+              <th>时长</th>
+              <th>句子</th>
+              <th>大小</th>
+              <th>更新时间</th>
+              <th class="actions-column">操作</th>
+            </tr>
+          </thead>
+          <tbody id="video-rows"></tbody>
+        </table>
+      </div>
+      <div id="empty-state" class="empty-state" hidden>暂无视频</div>
+    </section>
+  </main>
+
+  <dialog id="sentence-dialog">
+    <div class="dialog-heading">
+      <div>
+        <h2 id="sentence-title">句子文本</h2>
+        <span id="sentence-meta"></span>
+      </div>
+      <button id="close-dialog" class="icon-button" type="button" aria-label="关闭">X</button>
+    </div>
+    <div class="sentence-scroll">
+      <table>
+        <thead>
+          <tr><th>#</th><th>时间</th><th>文本</th><th>语言</th><th>操作</th></tr>
+        </thead>
+        <tbody id="sentence-rows"></tbody>
+      </table>
+    </div>
+  </dialog>
+
+  <script src="/static/admin.js" defer></script>
+</body>
+</html>
diff --git a/sentence_api/static/admin.js b/sentence_api/static/admin.js
new file mode 100644
index 0000000..90db333
--- /dev/null
+++ b/sentence_api/static/admin.js
@@ -0,0 +1,282 @@
+const state = {
+  adminKey: localStorage.getItem("oralTrainerAdminKey") || "",
+  videos: [],
+};
+
+const elements = {
+  adminKey: document.querySelector("#admin-key"),
+  saveKey: document.querySelector("#save-key"),
+  refresh: document.querySelector("#refresh"),
+  serviceState: document.querySelector("#service-state"),
+  uploadForm: document.querySelector("#upload-form"),
+  uploadState: document.querySelector("#upload-state"),
+  uploadProgress: document.querySelector("#upload-progress"),
+  videoRows: document.querySelector("#video-rows"),
+  videoCount: document.querySelector("#video-count"),
+  emptyState: document.querySelector("#empty-state"),
+  dialog: document.querySelector("#sentence-dialog"),
+  sentenceTitle: document.querySelector("#sentence-title"),
+  sentenceMeta: document.querySelector("#sentence-meta"),
+  sentenceRows: document.querySelector("#sentence-rows"),
+  closeDialog: document.querySelector("#close-dialog"),
+};
+
+elements.adminKey.value = state.adminKey;
+elements.saveKey.addEventListener("click", () => {
+  state.adminKey = elements.adminKey.value;
+  localStorage.setItem("oralTrainerAdminKey", state.adminKey);
+  elements.serviceState.textContent = "密钥已保存";
+});
+elements.refresh.addEventListener("click", loadVideos);
+elements.closeDialog.addEventListener("click", () => elements.dialog.close());
+elements.uploadForm.addEventListener("submit", uploadVideo);
+
+async function api(path, options = {}) {
+  const headers = new Headers(options.headers || {});
+  if (state.adminKey) headers.set("X-Admin-Key", state.adminKey);
+  const response = await fetch(path, { ...options, headers });
+  if (!response.ok) {
+    let detail = `HTTP ${response.status}`;
+    try {
+      const payload = await response.json();
+      detail = typeof payload.detail === "string" ? payload.detail : JSON.stringify(payload.detail);
+    } catch (_) {}
+    throw new Error(detail);
+  }
+  return response.status === 204 ? null : response.json();
+}
+
+async function checkHealth() {
+  try {
+    const health = await api("/healthz");
+    elements.serviceState.textContent = health.moss_configured ? "服务正常 / MOSS 已连接" : "服务正常 / MOSS 未配置";
+  } catch (error) {
+    elements.serviceState.textContent = `服务异常: ${error.message}`;
+  }
+}
+
+async function loadVideos() {
+  try {
+    const payload = await api("/api/v1/videos");
+    state.videos = payload.videos;
+    renderVideos();
+  } catch (error) {
+    elements.serviceState.textContent = `加载失败: ${error.message}`;
+  }
+}
+
+function renderVideos() {
+  elements.videoRows.replaceChildren();
+  elements.videoCount.textContent = `${state.videos.length} 个视频`;
+  elements.emptyState.hidden = state.videos.length !== 0;
+  for (const video of state.videos) {
+    const row = document.createElement("tr");
+    row.append(
+      cellWithError(video.title, video.error_message),
+      statusCell(video.status),
+      textCell(formatDuration(video.duration_ms)),
+      textCell(String(video.sentence_count)),
+      textCell(formatBytes(video.size_bytes)),
+      textCell(formatDate(video.updated_at)),
+      actionCell(video),
+    );
+    elements.videoRows.append(row);
+  }
+}
+
+function textCell(value) {
+  const cell = document.createElement("td");
+  cell.textContent = value;
+  return cell;
+}
+
+function cellWithError(title, error) {
+  const cell = textCell(title);
+  if (error) {
+    const message = document.createElement("div");
+    message.className = "error-message";
+    message.textContent = error;
+    cell.append(message);
+  }
+  return cell;
+}
+
+function statusCell(status) {
+  const cell = document.createElement("td");
+  const label = document.createElement("span");
+  label.className = `status status-${status}`;
+  label.textContent = { uploaded: "已上传", processing: "处理中", ready: "可用", failed: "失败" }[status] || status;
+  cell.append(label);
+  return cell;
+}
+
+function actionCell(video) {
+  const cell = document.createElement("td");
+  const actions = document.createElement("div");
+  actions.className = "actions";
+  const play = document.createElement("a");
+  play.href = video.stream_url;
+  play.target = "_blank";
+  play.rel = "noopener";
+  play.textContent = "播放";
+  actions.append(play);
+  actions.append(button("句子", () => openSentences(video)));
+  actions.append(button("重新处理", () => reprocess(video.video_hash)));
+  const remove = button("删除", () => deleteVideo(video));
+  remove.className = "danger";
+  actions.append(remove);
+  cell.append(actions);
+  return cell;
+}
+
+function button(label, handler) {
+  const result = document.createElement("button");
+  result.type = "button";
+  result.textContent = label;
+  result.addEventListener("click", handler);
+  return result;
+}
+
+function uploadVideo(event) {
+  event.preventDefault();
+  const file = document.querySelector("#video-file").files[0];
+  if (!file) return;
+  const title = document.querySelector("#video-title").value;
+  const language = document.querySelector("#video-language").value;
+  const query = new URLSearchParams({ filename: file.name });
+  if (title) query.set("title", title);
+  if (language) query.set("language", language);
+  elements.uploadProgress.hidden = false;
+  elements.uploadProgress.value = 0;
+  elements.uploadState.textContent = "上传中";
+  const request = new XMLHttpRequest();
+  request.open("PUT", `/api/v1/admin/videos/raw?${query.toString()}`);
+  if (state.adminKey) request.setRequestHeader("X-Admin-Key", state.adminKey);
+  request.setRequestHeader("Content-Type", file.type || "application/octet-stream");
+  request.upload.addEventListener("progress", (progress) => {
+    if (progress.lengthComputable) elements.uploadProgress.value = progress.loaded / progress.total * 100;
+  });
+  request.addEventListener("load", async () => {
+    if (request.status >= 200 && request.status < 300) {
+      elements.uploadState.textContent = "上传完成，正在处理";
+      elements.uploadForm.reset();
+      await loadVideos();
+    } else {
+      elements.uploadState.textContent = `上传失败: ${readXhrError(request)}`;
+    }
+    elements.uploadProgress.hidden = true;
+  });
+  request.addEventListener("error", () => {
+    elements.uploadState.textContent = "上传连接中断";
+    elements.uploadProgress.hidden = true;
+  });
+  request.send(file);
+}
+
+function readXhrError(request) {
+  try { return JSON.parse(request.responseText).detail || `HTTP ${request.status}`; }
+  catch (_) { return `HTTP ${request.status}`; }
+}
+
+async function openSentences(video) {
+  try {
+    const detail = await api(`/api/v1/videos/${video.video_hash}`);
+    elements.sentenceTitle.textContent = detail.video.title;
+    const sentences = detail.boundaries?.sentences || [];
+    elements.sentenceMeta.textContent = `${sentences.length} 句 / ${detail.video.video_hash}`;
+    elements.sentenceRows.replaceChildren();
+    for (const sentence of sentences) {
+      elements.sentenceRows.append(sentenceRow(video.video_hash, sentence));
+    }
+    elements.dialog.showModal();
+  } catch (error) {
+    elements.serviceState.textContent = `句子加载失败: ${error.message}`;
+  }
+}
+
+function sentenceRow(videoHash, sentence) {
+  const row = document.createElement("tr");
+  const textArea = document.createElement("textarea");
+  textArea.value = sentence.text || "";
+  const language = document.createElement("select");
+  for (const [value, label] of [["", "自动"], ["en", "英语"], ["zh", "中文"]]) {
+    const option = document.createElement("option");
+    option.value = value;
+    option.textContent = label;
+    option.selected = value === (sentence.language || "");
+    language.append(option);
+  }
+  const textContainer = document.createElement("td");
+  textContainer.append(textArea);
+  const languageContainer = document.createElement("td");
+  languageContainer.append(language);
+  const saveContainer = document.createElement("td");
+  const save = button("保存", async () => {
+    save.disabled = true;
+    try {
+      await api(`/api/v1/admin/videos/${videoHash}/sentences/${sentence.index}`, {
+        method: "PUT",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify({ text: textArea.value, language: language.value || null }),
+      });
+      save.textContent = "已保存";
+    } catch (error) {
+      save.textContent = error.message;
+    } finally {
+      window.setTimeout(() => { save.disabled = false; save.textContent = "保存"; }, 1600);
+    }
+  });
+  saveContainer.append(save);
+  row.append(
+    textCell(String(sentence.index + 1)),
+    textCell(`${formatDuration(sentence.start_ms)} - ${formatDuration(sentence.end_ms)}`),
+    textContainer,
+    languageContainer,
+    saveContainer,
+  );
+  return row;
+}
+
+async function reprocess(videoHash) {
+  try {
+    await api(`/api/v1/admin/videos/${videoHash}/process`, { method: "POST" });
+    await loadVideos();
+  } catch (error) {
+    elements.serviceState.textContent = `处理启动失败: ${error.message}`;
+  }
+}
+
+async function deleteVideo(video) {
+  if (!window.confirm(`删除“${video.title}”？`)) return;
+  try {
+    await api(`/api/v1/admin/videos/${video.video_hash}`, { method: "DELETE" });
+    await loadVideos();
+  } catch (error) {
+    elements.serviceState.textContent = `删除失败: ${error.message}`;
+  }
+}
+
+function formatDuration(ms) {
+  if (!ms) return "-";
+  const total = Math.floor(ms / 1000);
+  const hours = Math.floor(total / 3600);
+  const minutes = Math.floor(total % 3600 / 60);
+  const seconds = total % 60;
+  return hours ? `${hours}:${String(minutes).padStart(2, "0")}:${String(seconds).padStart(2, "0")}` : `${minutes}:${String(seconds).padStart(2, "0")}`;
+}
+
+function formatBytes(bytes) {
+  if (bytes < 1024 ** 2) return `${(bytes / 1024).toFixed(1)} KB`;
+  if (bytes < 1024 ** 3) return `${(bytes / 1024 ** 2).toFixed(1)} MB`;
+  return `${(bytes / 1024 ** 3).toFixed(2)} GB`;
+}
+
+function formatDate(value) {
+  return new Intl.DateTimeFormat("zh-CN", { dateStyle: "short", timeStyle: "short" }).format(new Date(value));
+}
+
+checkHealth();
+loadVideos();
+window.setInterval(() => {
+  if (!elements.dialog.open) loadVideos();
+}, 5000);
diff --git a/sentence_api/store.py b/sentence_api/store.py
index b8dc23e..1a3e3a7 100644
--- a/sentence_api/store.py
+++ b/sentence_api/store.py
@@ -23,7 +23,11 @@ class BoundaryStore:
         self.reload()
 
     def reload(self) -> None:
-        raw = json.loads(self.index_path.read_text(encoding="utf-8"))
+        raw = (
+            json.loads(self.index_path.read_text(encoding="utf-8"))
+            if self.index_path.is_file()
+            else {"videos": {}}
+        )
         entries = raw.get("videos", raw)
         if not isinstance(entries, dict):
             raise ValueError("index JSON must contain a 'videos' object")
diff --git a/sentence_api/tests/test_api.py b/sentence_api/tests/test_api.py
index ca45491..49c6052 100644
--- a/sentence_api/tests/test_api.py
+++ b/sentence_api/tests/test_api.py
@@ -1,8 +1,11 @@
 import json
+from dataclasses import replace
 
 from fastapi.testclient import TestClient
 
 from sentence_api.main import create_app
+from sentence_api.config import Settings
+from sentence_api.repository import VideoRepository
 from sentence_api.store import BoundaryStore
 
 
@@ -26,7 +29,19 @@ def make_client(tmp_path):
         }),
         encoding="utf-8",
     )
-    return TestClient(create_app(BoundaryStore(index_path)))
+    settings = replace(
+        Settings.from_env(),
+        data_dir=tmp_path / "data",
+        legacy_boundaries_file=index_path,
+    )
+    settings.ensure_directories()
+    return TestClient(
+        create_app(
+            BoundaryStore(index_path),
+            settings=settings,
+            repository=VideoRepository(settings.database_path),
+        )
+    )
 
 
 def test_lookup_returns_boundaries(tmp_path):
diff --git a/sentence_api/tests/test_assessment_api.py b/sentence_api/tests/test_assessment_api.py
new file mode 100644
index 0000000..fcb19e6
--- /dev/null
+++ b/sentence_api/tests/test_assessment_api.py
@@ -0,0 +1,146 @@
+import io
+import json
+import wave
+from dataclasses import replace
+
+import numpy as np
+from fastapi.testclient import TestClient
+
+from sentence_api.config import Settings
+from sentence_api.main import create_app
+from sentence_api.models import SentenceBoundaryDocument
+from sentence_api.repository import VideoRepository
+from sentence_api.store import BoundaryStore
+from sentence_api.transcription import Transcript
+
+
+VIDEO_HASH = "c" * 64
+
+
+class FakeTranscriber:
+    available = True
+
+    def transcribe(self, audio_path, language=None):
+        return Transcript(text="The meeting starts at nine.", segments=[])
+
+
+def make_wav() -> bytes:
+    sample_rate = 16_000
+    time = np.arange(sample_rate * 2, dtype=np.float64) / sample_rate
+    samples = (np.sin(2 * np.pi * 220 * time) * 8000).astype("<i2")
+    output = io.BytesIO()
+    with wave.open(output, "wb") as wav_file:
+        wav_file.setnchannels(1)
+        wav_file.setsampwidth(2)
+        wav_file.setframerate(sample_rate)
+        wav_file.writeframes(samples.tobytes())
+    return output.getvalue()
+
+
+def make_client(tmp_path, client_api_key=""):
+    legacy_path = tmp_path / "legacy.json"
+    legacy_path.write_text(json.dumps({"videos": {}}), encoding="utf-8")
+    settings = replace(
+        Settings.from_env(),
+        data_dir=tmp_path / "data",
+        legacy_boundaries_file=legacy_path,
+        admin_api_key="test-admin-key",
+        client_api_key=client_api_key,
+    )
+    settings.ensure_directories()
+    repository = VideoRepository(settings.database_path)
+    stored_filename = f"{VIDEO_HASH}.mp4"
+    (settings.videos_dir / stored_filename).write_bytes(b"video-placeholder")
+    repository.upsert_upload(
+        video_hash=VIDEO_HASH,
+        title="Lesson",
+        filename="lesson.mp4",
+        stored_filename=stored_filename,
+        content_type="video/mp4",
+        size_bytes=17,
+        language="en",
+    )
+    repository.save_processing_result(
+        SentenceBoundaryDocument(
+            video_hash=VIDEO_HASH,
+            duration_ms=2000,
+            algorithm_version="test-v1",
+            sentences=[
+                {
+                    "index": 0,
+                    "start_ms": 0,
+                    "end_ms": 2000,
+                    "text": "The meeting starts at nine.",
+                    "language": "en",
+                    "reference_speech_duration_ms": 2000,
+                }
+            ],
+        ),
+        "The meeting starts at nine.",
+    )
+    app = create_app(
+        BoundaryStore(legacy_path),
+        settings=settings,
+        repository=repository,
+        transcriber=FakeTranscriber(),
+    )
+    return TestClient(app)
+
+
+def test_assessment_returns_duration_and_content_breakdown(tmp_path):
+    client = make_client(tmp_path)
+
+    response = client.post(
+        f"/api/v1/videos/{VIDEO_HASH}/sentences/0/assessments",
+        files={"audio": ("reading.wav", make_wav(), "audio/wav")},
+        data={"language": "en"},
+    )
+
+    assert response.status_code == 200, response.text
+    payload = response.json()
+    assert payload["overall_score"] == 100
+    assert payload["passed"] is True
+    assert 0.99 <= payload["duration_ratio"] <= 1.02
+    assert payload["pronunciation_score"] is None
+    assert payload["details"]["phoneme_scoring"] == "not_enabled"
+
+
+def test_admin_routes_require_key(tmp_path):
+    client = make_client(tmp_path)
+
+    response = client.delete(f"/api/v1/admin/videos/{VIDEO_HASH}")
+
+    assert response.status_code == 401
+
+
+def test_catalog_and_range_streaming(tmp_path):
+    client = make_client(tmp_path)
+
+    catalog = client.get("/api/v1/videos")
+    stream = client.get(
+        f"/api/v1/videos/{VIDEO_HASH}/content",
+        headers={"Range": "bytes=0-4"},
+    )
+
+    assert catalog.status_code == 200
+    assert catalog.json()["videos"][0]["stream_url"].endswith(f"/{VIDEO_HASH}/content")
+    assert stream.status_code == 206
+    assert stream.content == b"video"
+
+
+def test_assessment_client_key_is_enforced_when_configured(tmp_path):
+    client = make_client(tmp_path, client_api_key="tablet-key")
+    endpoint = f"/api/v1/videos/{VIDEO_HASH}/sentences/0/assessments"
+
+    unauthorized = client.post(
+        endpoint,
+        files={"audio": ("reading.wav", make_wav(), "audio/wav")},
+    )
+    authorized = client.post(
+        endpoint,
+        headers={"X-Client-Key": "tablet-key"},
+        files={"audio": ("reading.wav", make_wav(), "audio/wav")},
+    )
+
+    assert unauthorized.status_code == 401
+    assert authorized.status_code == 200
diff --git a/sentence_api/tests/test_scoring.py b/sentence_api/tests/test_scoring.py
new file mode 100644
index 0000000..11b9ed5
--- /dev/null
+++ b/sentence_api/tests/test_scoring.py
@@ -0,0 +1,56 @@
+import numpy as np
+
+from sentence_api.audio_metrics import AudioMetrics, analyze_samples
+from sentence_api.scoring import align_text, duration_similarity_score, score_attempt, tokenize
+
+
+def test_tokenize_supports_mixed_chinese_and_english():
+    assert tokenize("你好，World! Don't stop.") == ["你", "好", "world", "don't", "stop"]
+
+
+def test_identical_reading_with_matching_duration_scores_100():
+    result = score_attempt(
+        reference_text="The meeting starts at nine.",
+        recognized_text="The meeting starts at nine",
+        reference_speech_duration_ms=2000,
+        student_metrics=AudioMetrics(
+            recording_duration_ms=2400,
+            speech_duration_ms=2000,
+            internal_silence_ms=0,
+            internal_pause_ratio=0.0,
+        ),
+    )
+
+    assert result.overall_score == 100
+    assert result.duration_score == 100
+    assert result.missing_tokens == []
+
+
+def test_alignment_reports_missing_extra_and_substituted_tokens():
+    result = align_text(
+        "The meeting starts at nine",
+        "The lesson start at nine today",
+    )
+
+    assert result.content_score < 70
+    assert result.extra_tokens == ["today"]
+    assert ("meeting", "lesson") in result.substitutions
+    assert ("starts", "start") in result.substitutions
+
+
+def test_duration_score_allows_students_to_read_more_slowly():
+    assert duration_similarity_score(0.8) == 100
+    assert duration_similarity_score(1.3) == 100
+    assert duration_similarity_score(1.5) == 60
+    assert duration_similarity_score(1.8) == 20
+
+
+def test_vad_excludes_leading_and_trailing_silence():
+    sample_rate = 16_000
+    silence = np.zeros(sample_rate // 2, dtype=np.float32)
+    time = np.arange(sample_rate, dtype=np.float32) / sample_rate
+    speech = (0.25 * np.sin(2 * np.pi * 220 * time)).astype(np.float32)
+    metrics = analyze_samples(np.concatenate([silence, speech, silence]), sample_rate)
+
+    assert metrics.recording_duration_ms == 2000
+    assert 900 <= metrics.speech_duration_ms <= 1050
diff --git a/sentence_api/tests/test_store.py b/sentence_api/tests/test_store.py
new file mode 100644
index 0000000..d0b330d
--- /dev/null
+++ b/sentence_api/tests/test_store.py
@@ -0,0 +1,7 @@
+from sentence_api.store import BoundaryStore
+
+
+def test_missing_legacy_index_starts_empty(tmp_path):
+    store = BoundaryStore(tmp_path / "not-created-yet.json")
+
+    assert store.count() == 0
diff --git a/sentence_api/transcription.py b/sentence_api/transcription.py
new file mode 100644
index 0000000..5267dea
--- /dev/null
+++ b/sentence_api/transcription.py
@@ -0,0 +1,134 @@
+import mimetypes
+import re
+from dataclasses import dataclass
+from pathlib import Path
+from typing import List, Optional, Protocol
+
+import httpx
+
+
+@dataclass(frozen=True)
+class TranscriptionSegment:
+    start_seconds: float
+    end_seconds: float
+    text: str
+    speaker: Optional[str] = None
+
+
+@dataclass(frozen=True)
+class Transcript:
+    text: str
+    segments: List[TranscriptionSegment]
+
+
+class Transcriber(Protocol):
+    @property
+    def available(self) -> bool:
+        ...
+
+    def transcribe(self, audio_path: Path, language: Optional[str] = None) -> Transcript:
+        ...
+
+
+class MossTranscriber:
+    def __init__(
+        self,
+        endpoint: str,
+        model: str,
+        timeout_seconds: float = 1800,
+        max_new_tokens: int = 65536,
+    ):
+        self.endpoint = endpoint.strip()
+        self.model = model
+        self.timeout_seconds = timeout_seconds
+        self.max_new_tokens = max_new_tokens
+
+    @property
+    def available(self) -> bool:
+        return bool(self.endpoint)
+
+    def transcribe(self, audio_path: Path, language: Optional[str] = None) -> Transcript:
+        if not self.available:
+            raise RuntimeError("MOSS transcription is not configured on this server.")
+        content_type = mimetypes.guess_type(audio_path.name)[0] or "application/octet-stream"
+        data = {
+            "model": self.model,
+            "response_format": "verbose_json",
+            "temperature": "0",
+            "max_new_tokens": str(self.max_new_tokens),
+        }
+        if language:
+            data["language"] = language
+
+        try:
+            with audio_path.open("rb") as audio_file:
+                response = httpx.post(
+                    self.endpoint,
+                    data=data,
+                    files={"file": (audio_path.name, audio_file, content_type)},
+                    timeout=httpx.Timeout(self.timeout_seconds, connect=30),
+                )
+            response.raise_for_status()
+            payload = response.json()
+        except (httpx.HTTPError, ValueError) as exc:
+            raise RuntimeError(f"MOSS transcription request failed: {exc}") from exc
+
+        raw_text = str(payload.get("text") or "").strip()
+        segments = _parse_json_segments(payload.get("segments"))
+        if not segments:
+            segments = _parse_compact_segments(raw_text)
+        plain_text = " ".join(segment.text for segment in segments).strip() or raw_text
+        return Transcript(text=plain_text, segments=segments)
+
+
+def _parse_json_segments(raw_segments) -> List[TranscriptionSegment]:
+    if not isinstance(raw_segments, list):
+        return []
+    segments = []
+    for item in raw_segments:
+        if not isinstance(item, dict):
+            continue
+        try:
+            start = float(item["start"])
+            end = float(item["end"])
+        except (KeyError, TypeError, ValueError):
+            continue
+        text = str(item.get("text") or "").strip()
+        if text and end > start >= 0:
+            speaker = item.get("speaker")
+            segments.append(
+                TranscriptionSegment(
+                    start_seconds=start,
+                    end_seconds=end,
+                    text=text,
+                    speaker=str(speaker) if speaker is not None else None,
+                )
+            )
+    return segments
+
+
+_COMPACT_SEGMENT = re.compile(
+    r"\[(?P<start>\d+(?:\.\d+)?)\]"
+    r"\[(?P<speaker>S\d+)\]"
+    r"(?P<text>.*?)"
+    r"\[(?P<end>\d+(?:\.\d+)?)\]",
+    flags=re.DOTALL,
+)
+
+
+def _parse_compact_segments(text: str) -> List[TranscriptionSegment]:
+    segments = []
+    for match in _COMPACT_SEGMENT.finditer(text):
+        start = float(match.group("start"))
+        end = float(match.group("end"))
+        segment_text = match.group("text").strip()
+        if segment_text and end > start:
+            segments.append(
+                TranscriptionSegment(
+                    start_seconds=start,
+                    end_seconds=end,
+                    text=segment_text,
+                    speaker=match.group("speaker"),
+                )
+            )
+    return segments