fixed a gain

sentence_api/DEPLOYMENT.md

@@ -159,7 +159,8 @@ segment 会自动合并成同一句。若转写服务支持词级时间戳
 标点所在单词的时间戳；不支持时自动按文字长度在 segment 内估算，请求失败会回退为
 不带词级时间戳的普通请求。MOSS 不支持该参数时同样自动回退，不影响转写。
 每句结束时间默认再向后顺延 300ms（`MOSS_END_PADDING_MS`，可在 200–500 之间调整），
-避免句子末尾发音被切掉；顺延不会越过下一句的开头。
+避免句子末尾发音被切掉；顺延不会越过下一句的开头。顺延后还会用音频检测句子间的
+停顿，把结束时间拉回到下一句语音真正开始之前，避免偶尔切到下一句的开头。
 
 ## 3. 部署 API
 

sentence_api/audio_metrics.py

@@ -7,6 +7,10 @@ import numpy as np
 
 SAMPLE_RATE = 16_000
 FRAME_SAMPLES = 480
+# A silence run at least this long (in 30 ms frames) separates two sentences.
+# speech_frame_mask bridges gaps of up to 5 frames, so use 6 frames (~180 ms).
+MIN_PAUSE_FRAMES = 6
+ONSET_MARGIN_MS = 50
 
 
 class AudioAnalysisError(RuntimeError):
@@ -108,6 +112,44 @@ def speech_frame_mask(samples: np.ndarray, sample_rate: int = SAMPLE_RATE) -> np
     return speech
 
 
+def refine_sentence_end_ms(
+    samples: np.ndarray,
+    sample_rate: int = SAMPLE_RATE,
+    *,
+    raw_end_ms: int,
+    padded_end_ms: int,
+) -> int:
+    """Pull a padded sentence end back to just before the next sentence's speech.
+
+    Whisper's timestamps are not always aligned with the real audio: the next
+    segment's start can be later than the actual speech onset, so a fixed
+    end-padding may occasionally run into the next sentence's beginning. This
+    finds the first silence run of at least MIN_PAUSE_FRAMES inside the padded
+    region and stops the sentence just before the speech that follows it.
+    """
+    if padded_end_ms <= raw_end_ms:
+        return padded_end_ms
+    frame_ms = 1000 * max(1, int(round(sample_rate * 0.03))) / sample_rate
+    start_sample = max(0, int(raw_end_ms / 1000 * sample_rate))
+    end_sample = min(samples.size, int(padded_end_ms / 1000 * sample_rate))
+    if end_sample <= start_sample:
+        return padded_end_ms
+    try:
+        speech = speech_frame_mask(samples[start_sample:end_sample], sample_rate)
+    except AudioAnalysisError:
+        return padded_end_ms
+    silence_frames = 0
+    for index, is_speech in enumerate(speech):
+        if not is_speech:
+            silence_frames += 1
+            continue
+        if silence_frames >= MIN_PAUSE_FRAMES:
+            onset_ms = int(raw_end_ms + index * frame_ms)
+            return max(raw_end_ms, min(padded_end_ms, onset_ms - ONSET_MARGIN_MS))
+        silence_frames = 0
+    return padded_end_ms
+
+
 def _bridge_false_runs(values: np.ndarray, max_frames: int) -> None:
     start = None
     for index, value in enumerate(values):

sentence_api/processing.py

@@ -4,7 +4,12 @@ import uuid
 from pathlib import Path
 from typing import List, Optional
 
-from .audio_metrics import AudioAnalysisError, analyze_samples, decode_audio_mono
+from .audio_metrics import (
+    AudioAnalysisError,
+    analyze_samples,
+    decode_audio_mono,
+    refine_sentence_end_ms,
+)
 from .config import Settings
 from .generate_boundaries import ALGORITHM_VERSION, make_entry
 from .models import SentenceBoundary, SentenceBoundaryDocument
@@ -112,15 +117,19 @@ def document_from_transcript(
     sentence_segments = split_sentences_at_punctuation(transcript.segments)
     for index, sentence_segment in enumerate(sentence_segments):
         start_ms = max(previous_end, int(round(sentence_segment.start_seconds * 1000)))
+        raw_end_ms = int(round(sentence_segment.end_seconds * 1000))
         end_ms = min(
             duration_ms,
-            int(round(sentence_segment.end_seconds * 1000)) + end_padding_ms,
+            raw_end_ms + end_padding_ms,
         )
         if index + 1 < len(sentence_segments):
             next_start_ms = int(
                 round(sentence_segments[index + 1].start_seconds * 1000)
             )
             end_ms = min(end_ms, next_start_ms)
+        end_ms = refine_sentence_end_ms(
+            samples, sample_rate, raw_end_ms=raw_end_ms, padded_end_ms=end_ms
+        )
         if not sentence_segment.text.strip() or end_ms <= start_ms:
             continue
         start_sample = max(0, int(start_ms / 1000 * sample_rate))