update:所有ASR均改为PCM直接识别，通过delete_audio参数判断要不要储存wav

main/xiaozhi-server/core/providers/asr/base.py

@@ -9,8 +9,8 @@ logger = setup_logging()
 
 class ASRProviderBase(ABC):
     @abstractmethod
-    def save_audio_to_file(self, opus_data: List[bytes], session_id: str) -> str:
+    def save_audio_to_file(self, pcm_data: List[bytes], session_id: str) -> str:
-        """解码Opus数据并保存为WAV文件"""
+        """PCM数据保存为WAV文件"""
         pass
 
     @abstractmethod

main/xiaozhi-server/core/providers/asr/doubao.py

@@ -89,6 +89,7 @@ class ASRProvider(ASRProviderBase):
         self.cluster = config.get("cluster")
         self.access_token = config.get("access_token")
         self.output_dir = config.get("output_dir")
+        self.delete_audio_file = delete_audio_file
 
         self.host = "openspeech.bytedance.com"
         self.ws_url = f"wss://{self.host}/api/v2/asr"
@@ -98,21 +99,11 @@ class ASRProvider(ASRProviderBase):
         # 确保输出目录存在
         os.makedirs(self.output_dir, exist_ok=True)
 
-    def save_audio_to_file(self, opus_data: List[bytes], session_id: str) -> str:
+    def save_audio_to_file(self, pcm_data: List[bytes], session_id: str) -> str:
-        """将Opus音频数据解码并保存为WAV文件"""
+        """PCM数据保存为WAV文件"""
         file_name = f"asr_{session_id}_{uuid.uuid4()}.wav"
         file_path = os.path.join(self.output_dir, file_name)
 
-        decoder = opuslib_next.Decoder(16000, 1)  # 16kHz, 单声道
-        pcm_data = []
-
-        for opus_packet in opus_data:
-            try:
-                pcm_frame = decoder.decode(opus_packet, 960)  # 960 samples = 60ms
-                pcm_data.append(pcm_frame)
-            except opuslib_next.OpusError as e:
-                logger.bind(tag=TAG).error(f"Opus解码错误: {e}", exc_info=True)
-
         with wave.open(file_path, "wb") as wf:
             wf.setnchannels(1)
             wf.setsampwidth(2)  # 2 bytes = 16-bit
@@ -274,6 +265,12 @@ class ASRProvider(ASRProviderBase):
             pcm_data = self.decode_opus(opus_data, session_id)
             combined_pcm_data = b"".join(pcm_data)
 
+            # 判断是否保存为WAV文件
+            if self.delete_audio_file:
+                pass
+            else:
+                self.save_audio_to_file(pcm_data, session_id)
+
             # 直接使用PCM数据
             # 计算分段大小 (单声道, 16bit, 16kHz采样率)
             size_per_sec = 1 * 2 * 16000  # nchannels * sampwidth * framerate

main/xiaozhi-server/core/providers/asr/fun_local.py

@@ -51,7 +51,7 @@ class ASRProvider(ASRProviderBase):
             )
 
     def save_audio_to_file(self, opus_data: List[bytes], session_id: str) -> str:
-        """将Opus音频数据解码并保存为WAV文件"""
+        """PCM数据保存为WAV文件"""
         file_name = f"asr_{session_id}_{uuid.uuid4()}.wav"
         file_path = os.path.join(self.output_dir, file_name)
 
@@ -72,20 +72,40 @@ class ASRProvider(ASRProviderBase):
             wf.writeframes(b"".join(pcm_data))
 
         return file_path
+    
+    @staticmethod
+    def decode_opus(opus_data: List[bytes], session_id: str) -> List[bytes]:
+
+        decoder = opuslib_next.Decoder(16000, 1)  # 16kHz, 单声道
+        pcm_data = []
+
+        for opus_packet in opus_data:
+            try:
+                pcm_frame = decoder.decode(opus_packet, 960)  # 960 samples = 60ms
+                pcm_data.append(pcm_frame)
+            except opuslib_next.OpusError as e:
+                logger.bind(tag=TAG).error(f"Opus解码错误: {e}", exc_info=True)
+
+        return pcm_data
 
     async def speech_to_text(self, opus_data: List[bytes], session_id: str) -> Tuple[Optional[str], Optional[str]]:
         """语音转文本主处理逻辑"""
         file_path = None
         try:
-            # 保存音频文件
+            # 合并所有opus数据包
-            start_time = time.time()
+            pcm_data = self.decode_opus(opus_data, session_id)
-            file_path = self.save_audio_to_file(opus_data, session_id)
+            combined_pcm_data = b"".join(pcm_data)
-            logger.bind(tag=TAG).debug(f"音频文件保存耗时: {time.time() - start_time:.3f}s | 路径: {file_path}")
+
+            # 判断是否保存为WAV文件
+            if self.delete_audio_file:
+                pass
+            else:
+                self.save_audio_to_file(pcm_data, session_id)
 
             # 语音识别
             start_time = time.time()
             result = self.model.generate(
-                input=file_path,
+                input=combined_pcm_data,
                 cache={},
                 language="auto",
                 use_itn=True,

main/xiaozhi-server/core/providers/asr/sherpa_onnx_local.py

@@ -84,10 +84,21 @@ class ASRProvider(ASRProviderBase):
             )
 
     def save_audio_to_file(self, opus_data: List[bytes], session_id: str) -> str:
-        """将Opus音频数据解码并保存为WAV文件"""
+        """PCM数据保存为WAV文件"""
         file_name = f"asr_{session_id}_{uuid.uuid4()}.wav"
         file_path = os.path.join(self.output_dir, file_name)
 
+        with wave.open(file_path, "wb") as wf:
+            wf.setnchannels(1)
+            wf.setsampwidth(2)  # 2 bytes = 16-bit
+            wf.setframerate(16000)
+            wf.writeframes(b"".join(pcm_data))
+
+        return file_path
+
+    @staticmethod
+    def decode_opus(opus_data: List[bytes], session_id: str) -> List[bytes]:
+
         decoder = opuslib_next.Decoder(16000, 1)  # 16kHz, 单声道
         pcm_data = []
 
@@ -98,13 +109,7 @@ class ASRProvider(ASRProviderBase):
             except opuslib_next.OpusError as e:
                 logger.bind(tag=TAG).error(f"Opus解码错误: {e}", exc_info=True)
 
-        with wave.open(file_path, "wb") as wf:
+        return pcm_data
-            wf.setnchannels(1)
-            wf.setsampwidth(2)  # 2 bytes = 16-bit
-            wf.setframerate(16000)
-            wf.writeframes(b"".join(pcm_data))
-
-        return file_path
 
     def read_wave(self, wave_filename: str) -> Tuple[np.ndarray, int]:
         """
@@ -136,7 +141,8 @@ class ASRProvider(ASRProviderBase):
         try:
             # 保存音频文件
             start_time = time.time()
-            file_path = self.save_audio_to_file(opus_data, session_id)
+            pcm_data = self.decode_opus(opus_data, session_id)
+            file_path = self.save_audio_to_file(pcm_data, session_id)
             logger.bind(tag=TAG).debug(f"音频文件保存耗时: {time.time() - start_time:.3f}s | 路径: {file_path}")
 
             # 语音识别

main/xiaozhi-server/core/providers/asr/tencent.py

@@ -31,21 +31,11 @@ class ASRProvider(ASRProviderBase):
         os.makedirs(self.output_dir, exist_ok=True)
 
     def save_audio_to_file(self, opus_data: List[bytes], session_id: str) -> str:
-        """将Opus音频数据解码并保存为WAV文件"""
+        """PCM数据保存为WAV文件"""
 
         file_name = f"tencent_asr_{session_id}_{uuid.uuid4()}.wav"
         file_path = os.path.join(self.output_dir, file_name)
         
-        decoder = opuslib_next.Decoder(16000, 1)  # 16kHz, 单声道
-        pcm_data = []
-        
-        for opus_packet in opus_data:
-            try:
-                pcm_frame = decoder.decode(opus_packet, 960)  # 960 samples = 60ms
-                pcm_data.append(pcm_frame)
-            except opuslib_next.OpusError as e:
-                logger.bind(tag=TAG).error(f"Opus解码错误: {e}", exc_info=True)
-        
         with wave.open(file_path, "wb") as wf:
             wf.setnchannels(1)
             wf.setsampwidth(2)  # 2 bytes = 16-bit
@@ -85,6 +75,12 @@ class ASRProvider(ASRProviderBase):
 
             # 将Opus音频数据解码为PCM
             pcm_data = self.decode_opus(opus_data)
+
+            # 判断是否保存为WAV文件
+            if self.delete_audio_file:
+                pass
+            else:
+                self.save_audio_to_file(pcm_data, session_id)
             
             # 将音频数据转换为Base64编码
             base64_audio = base64.b64encode(pcm_data).decode('utf-8')