add translate from 7b model when partial==False

transcribe/translatepipes.py

@@ -1,4 +1,4 @@
-from transcribe.pipelines import WhisperPipe, MetaItem, WhisperChinese, Translate7BPipe, FunASRPipe, VadPipe
+from transcribe.pipelines import WhisperPipe, MetaItem, WhisperChinese, Translate7BPipe, FunASRPipe, VadPipe, TranslatePipe
 
 
 class TranslatePipes:
@@ -11,7 +11,7 @@ class TranslatePipes:
         self._funasr_pipe = self._launch_process(FunASRPipe())
 
         # llm 翻译
-        # self._translate_pipe = self._launch_process(TranslatePipe())
+        self._translate_pipe = self._launch_process(TranslatePipe())
         self._translate_7b_pipe = self._launch_process(Translate7BPipe())
         # vad 
         self._vad_pipe = self._launch_process(VadPipe())

transcribe/whisper_llm_serve.py

@@ -68,7 +68,7 @@ class WhisperTranscriptionService:
         self._frame_processing_thread_stop = threading.Event()
 
         self.translate_thread = _start_thread(self._transcription_processing_loop)
-        self.frame_processing_thread = _start_thread(self._frame_processing_loop)
+        self.frame_processing_thread = _start_thread(self._read_frame_processing_loop)
         self.row_number = 0
         # for test
         self._transcribe_time_cost = 0.
@@ -98,7 +98,7 @@ class WhisperTranscriptionService:
         return speech_audio, speech_status
 
 
-    def _frame_processing_loop(self) -> None:
+    def _read_frame_processing_loop(self) -> None:
         """从队列获取音频帧并合并到缓冲区"""
         while not self._frame_processing_thread_stop.is_set():
             try:
@@ -109,17 +109,21 @@ class WhisperTranscriptionService:
                     continue
 
                 with self.lock:
+                    self.frames_np = np.append(self.frames_np, frame_np)
+
+                     # 音频开始时间节点 用来统计时间来 达到最小断句时间长度
                     if speech_status == "START" and self.frames_np_start_timestamp is None:
                         self.frames_np_start_timestamp = time.time()
-                    # 添加音频到音频缓冲区
-                    self.frames_np = np.append(self.frames_np, frame_np)
+
+                    # 音频最长时间缓冲区限制，超过了就强制断句
                     if len(self.frames_np) >= self.sample_rate * config.MAX_SPEECH_DURATION_S:
                         audio_array=self.frames_np.copy()
                         self.full_segments_queue.appendleft(audio_array) # 根据时间是否满足三秒长度 来整合音频块
                         self.frames_np_start_timestamp = time.time()
                         self.frames_np = np.array([], dtype=np.float32)
-
-                    elif speech_status == "END" and len(self.frames_np) > 0 and self.frames_np_start_timestamp:
+                    
+                    # 音频结束信号的时候 整合当前缓冲区
+                    elif speech_status == "END" and len(self.frames_np) > 0:
                         time_diff = time.time() - self.frames_np_start_timestamp
                         if time_diff >= config.FRAME_SCOPE_TIME_THRESHOLD:
                             audio_array=self.frames_np.copy()
@@ -127,7 +131,7 @@ class WhisperTranscriptionService:
                             self.frames_np_start_timestamp = None
                             self.frames_np = np.array([], dtype=np.float32)
                         else:
-                            logger.debug(f"🥳 当前时间与上一句的时间差: {time_diff:.2f}s,继续增加缓冲区")
+                            logger.debug(f"🥳 当前时间与上一句的时间差: {time_diff:.2f}s,继续保留在缓冲区")
                     
             except queue.Empty:
                 pass
@@ -139,7 +143,7 @@ class WhisperTranscriptionService:
         while not self._translate_thread_stop.is_set():
             
             if len(self.frames_np) ==0:
-                time.sleep(0.01)
+                time.sleep(0.1)
                 continue
 
             with self.lock:
@@ -163,21 +167,24 @@ class WhisperTranscriptionService:
 
             if len(segments):
                 seg_text = self.text_separator.join(seg.text for seg in segments)
+                # 整行
+                if not partial:
+                    translated_content = self._translate_text(seg_text)
+                    self.row_number += 1
+                    frame_epoch = 1
+                else:
+                    translated_content = self._translate_text_large(seg_text)
+                    frame_epoch += 1
+
                 result = TransResult(
                     seg_id=self.row_number,
                     context=seg_text,
                     from_=self.source_language,
                     to=self.target_language,
-                    tran_content=self._translate_text_large(seg_text),
+                    tran_content=translated_content,
                     partial=partial
                 )
                 self._send_result_to_client(result)
-                if not partial:
-                    self.row_number += 1
-                    frame_epoch = 1
-                else:
-                    frame_epoch += 1
-
     
 
 
@@ -221,6 +228,7 @@ class WhisperTranscriptionService:
 
         result = self._translate_pipe.translate_large(text, self.source_language, self.target_language)
         translated_text = result.translate_content
+        
         time_diff = (time.perf_counter() - start_time)
         log_block("Translation large model time ", f"{time_diff:.3f}", "s")
         log_block("Translation large model output", f"{translated_text}")