diff --git a/sentence_api/static/admin.html b/sentence_api/static/admin.html
index fc28a24..78132ec 100644
--- a/sentence_api/static/admin.html
+++ b/sentence_api/static/admin.html
@@ -139,6 +139,6 @@
     </div>
   </dialog>
 
-  <script src="/static/admin.js?v=admin-sentence-batch-v3" defer></script>
+  <script src="/static/admin.js?v=admin-sentence-batch-v5" defer></script>
 </body>
 </html>
diff --git a/sentence_api/static/admin.js b/sentence_api/static/admin.js
index 1cfb442..4f2b368 100644
--- a/sentence_api/static/admin.js
+++ b/sentence_api/static/admin.js
@@ -389,13 +389,11 @@ async function handleSentenceDialogClose() {
   if (state.sentencePending) {
     const persisted = await saveSentenceDraft({ processTimings: false });
     if (!persisted) {
-      if (state.sentenceVideo && !elements.dialog.open) elements.dialog.showModal();
       return;
     }
   }
   const processed = await saveSentenceDraft({ processTimings: true });
   if (!processed) {
-    if (state.sentenceVideo && !elements.dialog.open) elements.dialog.showModal();
     return;
   }
   resetSentenceDraft();
@@ -417,8 +415,17 @@ async function saveSentenceDraft({ processTimings = false } = {}) {
         index,
         originStartMs,
         originEndMs,
+        sourceStartMs,
+        sourceEndMs,
         ...sentence
-      }) => sentence),
+      }) => ({
+        ...sentence,
+        ...(sourceStartMs !== undefined && sourceEndMs !== undefined ? {
+          source_start_ms: sourceStartMs,
+          source_end_ms: sourceEndMs,
+        } : {}),
+      }),
+      ),
     };
     const saved = await api(
       `/api/v1/admin/videos/${video.video_hash}/sentences/batch`,
diff --git a/sentence_api/tests/test_api.py b/sentence_api/tests/test_api.py
index 13d414a..e946237 100644
--- a/sentence_api/tests/test_api.py
+++ b/sentence_api/tests/test_api.py
@@ -936,7 +936,7 @@ def test_admin_batch_applies_merge_and_split_once(tmp_path, monkeypatch):
     assert all(not (item.zh_text or "").strip() for item in document.sentences)
 
 
-def test_admin_batch_can_persist_provisional_split_without_processing(tmp_path):
+def test_admin_batch_can_persist_provisional_split_without_processing(tmp_path, monkeypatch):
     settings = replace(Settings.from_env(), data_dir=tmp_path / "data")
     settings.ensure_directories()
     repository = VideoRepository(settings.database_path)
@@ -962,21 +962,43 @@ def test_admin_batch_can_persist_provisional_split_without_processing(tmp_path):
         None,
     )
     repository.save_sentence_translations(VIDEO_HASH, {1: "未变化。"})
+    (settings.videos_dir / f"{VIDEO_HASH}.mp4").write_bytes(b"video")
+    from sentence_api.transcription import Transcript, TranscriptionSegment, WordTimestamp
+
     calls = []
 
-    class MustNotTranscribe:
+    class DeferredTranscriber:
         available = True
+        allowed = False
 
         def transcribe(self, audio_path, language=None):
+            assert self.allowed, "provisional save must not transcribe"
             calls.append(audio_path)
-            raise AssertionError("provisional save must not transcribe")
+            return Transcript(
+                text="First part second part",
+                segments=[
+                    TranscriptionSegment(
+                        start_seconds=0.0,
+                        end_seconds=1.0,
+                        text="First part second part",
+                        words=[
+                            WordTimestamp(0.0, 0.2, "First"),
+                            WordTimestamp(0.2, 1.0, "part second part"),
+                        ],
+                    )
+                ],
+            )
 
+    monkeypatch.setattr(
+        "sentence_api.processing.extract_audio_segment",
+        lambda media_path, output_path, *, start_ms, duration_ms: output_path.write_bytes(b"audio"),
+    )
     client = TestClient(
         create_app(
             BoundaryStore(tmp_path / "boundaries.json"),
             settings=settings,
             repository=repository,
-            transcriber=MustNotTranscribe(),
+            transcriber=DeferredTranscriber(),
         )
     )
     response = client.put(
@@ -1001,6 +1023,25 @@ def test_admin_batch_can_persist_provisional_split_without_processing(tmp_path):
     assert [item.text for item in document.sentences] == ["First part", "second part", "Unchanged."]
     assert [item.zh_text for item in document.sentences] == [None, None, "未变化。"]
 
+    client.app.state.transcriber.allowed = True
+    processed = client.put(
+        f"/api/v1/admin/videos/{VIDEO_HASH}/sentences/batch",
+        json={
+            "process_timings": True,
+            "sentences": [
+                {"start_ms": 0, "end_ms": 2500, "text": "First part", "language": "en", "zh_text": None, "reference_speech_duration_ms": 2300, "source_start_ms": 0, "source_end_ms": 5000},
+                {"start_ms": 2500, "end_ms": 5000, "text": "second part", "language": "en", "zh_text": None, "reference_speech_duration_ms": 2300, "source_start_ms": 0, "source_end_ms": 5000},
+                {"start_ms": 5000, "end_ms": 12000, "text": "Unchanged.", "language": "en", "zh_text": "未变化。", "reference_speech_duration_ms": 6400},
+            ],
+        },
+    )
+
+    assert processed.status_code == 200, processed.text
+    processed_payload = processed.json()
+    assert [item["start_ms"] for item in processed_payload] == [0, 1000, 5000]
+    assert [item["end_ms"] for item in processed_payload] == [1000, 5000, 12000]
+    assert len(calls) == 1
+
 
 def test_admin_sentence_merge_does_not_start_subtitle_translation(tmp_path, monkeypatch):
     client = make_client(tmp_path)