fix bug of count word length

transcribe/pipelines/pipe_vad.py

@@ -41,7 +41,7 @@ class VadPipe(BasePipe):
     def process(self, in_data: MetaItem) -> MetaItem:
         source_audio = in_data.source_audio
         source_audio = np.frombuffer(source_audio, dtype=np.float32)
-        
+        # source_audio = self.reduce_noise(source_audio)
         send_audio = b""
         speech_timestamps = get_speech_timestamps(torch.Tensor(source_audio), self.model.silero_vad, sampling_rate=16000)
         

transcribe/strategy.py

@@ -149,11 +149,12 @@ class TranscriptBuffer:
 
     """
 
-    def __init__(self):
+    def __init__(self, separator):
         self._segments: List[str] = collections.deque(maxlen=2)     # 确认的完整段落
         self._sentences: List[str] = []    # 当前段落中的短句
         self._buffer: str = ""             # 当前缓冲中的文本
         self._current_seg_id: int = 0
+        self._separator = separator
 
     def get_seg_id(self) -> int:
         return self._current_seg_id
@@ -189,7 +190,8 @@ class TranscriptBuffer:
         if is_end_sentence:
             self.update_pending_text(stable_string)
             self.commit_line()
-            current_text_len = len(self.current_not_commit_text)
+            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)
             if current_text_len >=20:
                 self.commit_paragraph()
@@ -224,16 +226,18 @@ class TranscriptBuffer:
 
 
 class TranscriptStabilityAnalyzer:
-    def __init__(self) -> None:
-        self._transcript_buffer = TranscriptBuffer()
+    def __init__(self, separator) -> None:
+        self._transcript_buffer = TranscriptBuffer(separator=separator)
         self._transcript_history = TranscriptHistory()
+        self._separator = separator
+        logger.debug(f"Current separator: {self._separator}")
 
     def merge_chunks(self, chunks: List[TranscriptChunk])->str:
         return "".join(r.join() for r in chunks)
     
 
-    def analysis(self, separator, current: TranscriptChunk, buffer_duration: float) -> Iterator[TranscriptResult]:
-        current = TranscriptChunk(items=current, separator=separator)
+    def analysis(self, current: TranscriptChunk, buffer_duration: float) -> Iterator[TranscriptResult]:
+        current = TranscriptChunk(items=current, separator=self._separator)
         self._transcript_history.add(current)
 
         prev = self._transcript_history.previous_chunk()

transcribe/whisper_llm_serve.py

@@ -29,7 +29,7 @@ class WhisperTranscriptionService(ServeClientBase):
         self.target_language = dst_lang  # 目标翻译语言
         
         # 转录结果稳定性管理
-        self._transcrible_analysis = TranscriptStabilityAnalyzer()
+       
         self._translate_pipe = pipe
         
         # 音频处理相关
@@ -43,7 +43,7 @@ class WhisperTranscriptionService(ServeClientBase):
         
         # 发送就绪状态
         self.send_ready_state()
-        
+        self._transcrible_analysis = None
         # 启动处理线程
         self._translate_thread_stop = threading.Event()
         self._frame_processing_thread_stop = threading.Event()
@@ -75,6 +75,7 @@ class WhisperTranscriptionService(ServeClientBase):
         self.source_language = source_lang
         self.target_language = target_lang
         self.text_separator = self._get_text_separator(source_lang)
+        self._transcrible_analysis = TranscriptStabilityAnalyzer(self.text_separator)
 
     def add_audio_frames(self, frame_np: np.ndarray) -> None:
         """添加音频帧到处理队列"""
@@ -209,8 +210,7 @@ class WhisperTranscriptionService(ServeClientBase):
         if not segments:
             return
         
-        for ana_result in self._transcrible_analysis.analysis(
-            self.text_separator,segments, len(audio_buffer)/self.sample_rate):
+        for ana_result in self._transcrible_analysis.analysis(segments, len(audio_buffer)/self.sample_rate):
             if (cut_index :=ana_result.cut_index)>0:
                 # 更新音频缓冲区，移除已处理部分
                 self._update_audio_buffer(cut_index)