ajust time span

sentence_api/.env.example

@@ -14,3 +14,5 @@ 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
+# 每句结束时间向后顺延的毫秒数，避免句子末尾发音被切掉；按需在 200-500 之间调整
+MOSS_END_PADDING_MS=300

sentence_api/DEPLOYMENT.md

@@ -158,6 +158,8 @@ segment 会自动合并成同一句。若转写服务支持词级时间戳
 （`timestamp_granularities[]=word`，Whisper/Speaches 支持），句子的结束时间用
 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。
 
 ## 3. 部署 API
 

sentence_api/config.py

@@ -24,6 +24,7 @@ class Settings:
     moss_model: str
     moss_timeout_seconds: float
     moss_max_new_tokens: int
+    moss_end_padding_ms: int
     pass_score: float
 
     @classmethod
@@ -51,6 +52,7 @@ class Settings:
             ),
             moss_timeout_seconds=float(os.getenv("MOSS_TIMEOUT_SECONDS", "1800")),
             moss_max_new_tokens=int(os.getenv("MOSS_MAX_NEW_TOKENS", "65536")),
+            moss_end_padding_ms=int(os.getenv("MOSS_END_PADDING_MS", "300")),
             pass_score=float(os.getenv("ASSESSMENT_PASS_SCORE", "70")),
         )
 

sentence_api/processing.py

@@ -47,6 +47,7 @@ class VideoProcessor:
                     transcript=transcript,
                     language=video.get("language"),
                     audio_path=work_path,
+                    end_padding_ms=self.settings.moss_end_padding_ms,
                 )
                 if not document.sentences:
                     raise RuntimeError("MOSS returned no timestamped speech segments.")
@@ -103,13 +104,23 @@ def document_from_transcript(
     transcript: Transcript,
     language: Optional[str],
     audio_path: Path,
+    end_padding_ms: int = 300,
 ) -> SentenceBoundaryDocument:
     samples, sample_rate = decode_audio_mono(audio_path)
     sentences: List[SentenceBoundary] = []
     previous_end = 0
-    for sentence_segment in split_sentences_at_punctuation(transcript.segments):
+    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)))
-        end_ms = min(duration_ms, int(round(sentence_segment.end_seconds * 1000)))
+        end_ms = min(
+            duration_ms,
+            int(round(sentence_segment.end_seconds * 1000)) + 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)
         if not sentence_segment.text.strip() or end_ms <= start_ms:
             continue
         start_sample = max(0, int(start_ms / 1000 * sample_rate))