update and fix bug newline

transcribe/pipelines/pipe_vad.py

@@ -17,11 +17,11 @@ class VadPipe(BasePipe):
     def init(cls):
         if cls.model is None:
             cls.model = SileroVADProcessor(
-            activate_threshold=0.4,       # 降低以捕获更多音频
+            activate_threshold=0.45,       # 降低以捕获更多音频
             fusion_threshold=0.45,        # 提高以更好地融合语音片段
             min_speech_duration=0.2,      # 略微降低以捕获短音节
             max_speech_duration=20,       # 保持不变
-            # min_silence_duration=300,     # 增加到300毫秒，允许说话间的自然停顿
+            min_silence_duration=300,     # 增加到300毫秒，允许说话间的自然停顿
             sample_rate=cls.sample_rate # 采样率，音频信号的采样频率
         )
             cls.vac = FixedVADIterator(cls.model.silero_vad, sampling_rate=cls.sample_rate,)

transcribe/strategy.py

@@ -183,7 +183,7 @@ class TranscriptBuffer:
         """更新临时缓冲字符串"""
         self._buffer = text
 
-    def commit_line(self) -> None:
+    def commit_line(self,) -> None:
         """将缓冲字符串提交为短句"""
         if self._buffer:
             self._sentences.append(self._buffer)
@@ -219,6 +219,10 @@ class TranscriptBuffer:
         output = self.split_and_join(
                     text.replace(
                         self._separator, ""))
+        
+        logger.debug("==== rebuild string ====")
+        logger.debug(text)
+        logger.debug(output)
 
         return output
     
@@ -249,15 +253,18 @@ class TranscriptBuffer:
         return output
 
 
-    def update_and_commit(self, stable_string: str, remaining_string:str, is_end_sentence=False):
+    def update_and_commit(self, stable_strings: List[str], remaining_strings:List[str], is_end_sentence=False):
         if self.source_language == "en":
-            stable_string = self.rebuild(stable_string)
-            remaining_string = self.rebuild(remaining_string)
+            stable_strings = [self.rebuild(i) for i in stable_strings]
+            remaining_strings =[self.rebuild(i) for i in remaining_strings]
+        remaining_string = "".join(remaining_strings)
 
         logger.debug(f"{self.__dict__}")
         if is_end_sentence:
-            self.update_pending_text(stable_string)
-            self.commit_line()
+            for stable_str in stable_strings:
+                self.update_pending_text(stable_str)
+                self.commit_line()
+                
             current_text_len =  len(self.current_not_commit_text.split(self._separator)) if self._separator else len(self.current_not_commit_text)
             # current_text_len = len(self.current_not_commit_text.split(self._separator))
             self.update_pending_text(remaining_string)
@@ -266,8 +273,9 @@ class TranscriptBuffer:
                 self._current_seg_id += 1
                 return True
         else:
-            self.update_pending_text(stable_string)
-            self.commit_line()
+            for stable_str in stable_strings:
+                self.update_pending_text(stable_str)
+                self.commit_line()
             self.update_pending_text(remaining_string)
         return False
 
@@ -301,7 +309,7 @@ class TranscriptStabilityAnalyzer:
         logger.debug(f"Current separator: {self._separator}")
 
     def merge_chunks(self, chunks: List[TranscriptChunk])->str:
-        return "".join(r.join() for r in chunks)
+        return list(r.join() for r in chunks)
     
     
 
@@ -365,12 +373,12 @@ class TranscriptStabilityAnalyzer:
 
 
     def _yield_commit_results(self, stable_chunk, remaining_chunks, is_end_sentence: bool) -> Iterator[TranscriptResult]:
-        stable_str = stable_chunk.join() if hasattr(stable_chunk, "join") else self.merge_chunks(stable_chunk)
-        remaining_str = self.merge_chunks(remaining_chunks)
+        stable_str_list = [stable_chunk.join()] if hasattr(stable_chunk, "join") else self.merge_chunks(stable_chunk)
+        remaining_str_list = self.merge_chunks(remaining_chunks)
         frame_cut_index = stable_chunk[-1].get_buffer_index() if isinstance(stable_chunk, list) else stable_chunk.get_buffer_index()
         
         prev_seg_id = self._transcript_buffer.get_seg_id()
-        commit_paragraph = self._transcript_buffer.update_and_commit(stable_str, remaining_str, is_end_sentence)
+        commit_paragraph = self._transcript_buffer.update_and_commit(stable_str_list, remaining_str_list, is_end_sentence)
         logger.debug(f"current buffer: {self._transcript_buffer.__dict__}")
 
         if commit_paragraph:

transcribe/whisper_llm_serve.py

@@ -122,17 +122,17 @@ class WhisperTranscriptionService(ServeClientBase):
         frames = self.frames_np.copy()
 
         # 音频过短时的处理
-        if len(frames) <= 100:
+        if len(frames) <= 10:
             # 极短音频段，清空并返回None
-            self._update_audio_buffer(len(frames))
+            # self._update_audio_buffer(len(frames))
             return None
-        elif len(frames) < self.sample_rate:
+        if len(frames) < self.sample_rate:
             # 不足一秒的音频，补充静音
             silence_audio = np.zeros((self.sample_rate + 1000,), dtype=np.float32)
             silence_audio[-len(frames):] = frames
             return silence_audio.copy()
 
-        return frames.copy()
+        return frames
 
     def _transcribe_audio(self, audio_buffer: np.ndarray) -> List[TranscriptToken]:
         """转录音频并返回转录片段"""