feat: 服务端AEC功能

ref: 入口解码为pcm音频数据，vad和asr可直接使用

main/xiaozhi-server/core/connection.py

@@ -11,6 +11,8 @@ import threading
 import traceback
 import subprocess
 import websockets
+import opuslib_next
+import numpy as np
 
 from core.utils.util import (
     extract_json_from_string,
@@ -114,6 +116,7 @@ class ConnectionHandler:
         self.client_abort = False
         self.client_is_speaking = False
         self.client_listen_mode = "auto"
+        self.client_aec = False  # 是否启用了服务端AEC
 
         # 线程任务相关
         self.loop = None  # 在 handle_connection 中获取运行中的事件循环
@@ -153,7 +156,7 @@ class ConnectionHandler:
         # asr相关变量
         # 因为实际部署时可能会用到公共的本地ASR，不能把变量暴露给公共ASR
         # 所以涉及到ASR的变量，需要在这里定义，属于connection的私有变量
-        self.asr_audio = []
+        self.asr_audio = []  # 存储PCM帧列表，供VAD和ASR共享
         self.asr_audio_queue = queue.Queue()
         self.current_speaker = None  # 存储当前说话人
 
@@ -231,6 +234,9 @@ class ConnectionHandler:
             # 启动超时检查任务
             self.timeout_task = asyncio.create_task(self._check_timeout())
 
+            # 启动AEC缓存清理任务
+            self._aec_cache_cleanup_task = asyncio.create_task(self._check_aec_cache_expiry())
+
             self.welcome_msg = self.config["xiaozhi"]
             self.welcome_msg["session_id"] = self.session_id
 
@@ -366,12 +372,14 @@ class ConnectionHandler:
                 if handled:
                     return
 
-            # 不需要头部处理或没有头部时，直接处理原始消息
-            self.asr_audio_queue.put(message)
+            # 入口处直接解码PCM，避免VAD和ASR重复解码
+            pcm_frame = self._decode_opus_packet(message)
+            if pcm_frame:
+                self.asr_audio_queue.put(pcm_frame)
 
     async def _process_mqtt_audio_message(self, message):
         """
-        处理来自MQTT网关的音频消息，解析16字节头部并提取音频数据
+        处理来自MQTT网关的音频消息，解析16字节头部并提取音频数据，在入队前进行AEC处理
 
         Args:
             message: 包含头部的音频消息
@@ -380,58 +388,192 @@ class ConnectionHandler:
             bool: 是否成功处理了消息
         """
         try:
-            # 提取头部信息
+            # 解析timestamp
             timestamp = int.from_bytes(message[8:12], "big")
-            audio_length = int.from_bytes(message[12:16], "big")
 
-            # 提取音频数据
-            if audio_length > 0 and len(message) >= 16 + audio_length:
-                # 有指定长度，提取精确的音频数据
-                audio_data = message[16 : 16 + audio_length]
-                # 基于时间戳进行排序处理
-                self._process_websocket_audio(audio_data, timestamp)
-                return True
-            elif len(message) > 16:
-                # 没有指定长度或长度无效，去掉头部后处理剩余数据
-                audio_data = message[16:]
-                self.asr_audio_queue.put(audio_data)
+            audio_data = message[16:]
+            # 入口直接解码PCM
+            pcm_frame = self._decode_opus_packet(audio_data)
+            if not pcm_frame:
                 return True
+
+            # AEC处理：如果timestamp>0且启用了AEC
+            if timestamp > 0 and self.client_aec:
+                pcm_frame = self._apply_aec(timestamp, pcm_frame)
+
+            self.asr_audio_queue.put(pcm_frame)
+            return True
         except Exception as e:
             self.logger.bind(tag=TAG).error(f"解析WebSocket音频包失败: {e}")
 
         # 处理失败，返回False表示需要继续处理
         return False
 
-    def _process_websocket_audio(self, audio_data, timestamp):
-        """处理WebSocket格式的音频包"""
-        # 初始化时间戳序列管理
-        if not hasattr(self, "audio_timestamp_buffer"):
-            self.audio_timestamp_buffer = {}
-            self.last_processed_timestamp = 0
-            self.max_timestamp_buffer_size = 20
+    def _apply_aec(self, timestamp: int, pcm_frame: bytes) -> bytes:
+        """应用AEC处理 - 综合算法：互相关延迟估计 + Wiener滤波 + 频谱减法"""
+        try:
+            import numpy as np
 
-        # 如果时间戳是递增的，直接处理
-        if timestamp >= self.last_processed_timestamp:
-            self.asr_audio_queue.put(audio_data)
-            self.last_processed_timestamp = timestamp
+            if not pcm_frame or len(pcm_frame) == 0:
+                return pcm_frame
 
-            # 处理缓冲区中的后续包
-            processed_any = True
-            while processed_any:
-                processed_any = False
-                for ts in sorted(self.audio_timestamp_buffer.keys()):
-                    if ts > self.last_processed_timestamp:
-                        buffered_audio = self.audio_timestamp_buffer.pop(ts)
-                        self.asr_audio_queue.put(buffered_audio)
-                        self.last_processed_timestamp = ts
-                        processed_any = True
-                        break
-        else:
-            # 乱序包，暂存
-            if len(self.audio_timestamp_buffer) < self.max_timestamp_buffer_size:
-                self.audio_timestamp_buffer[timestamp] = audio_data
+            if not hasattr(self, "aec_audio_cache") or not self.aec_audio_cache:
+                return pcm_frame
+
+            mic_audio = np.frombuffer(pcm_frame, dtype=np.int16).astype(np.float32)
+            mic_rms = np.sqrt(np.mean(mic_audio ** 2))
+
+            if mic_rms < 100:
+                return pcm_frame
+
+            # 初始化
+            if not hasattr(self, "_aec_filter_state"):
+                self._aec_filter_state = np.zeros(512)  # 滤波器状态
+                self._aec_delay_ms = 200
+                self._aec_last_delay = 200
+                self._aec_coherence_history = []
+
+            sorted_timestamps = sorted(self.aec_audio_cache.keys())
+            if len(sorted_timestamps) < 2:
+                return pcm_frame
+
+            # ========== 匹配参考帧（对数功率谱匹配） ==========
+            n = len(mic_audio)
+            sorted_timestamps = sorted(self.aec_audio_cache.keys())
+
+            # 找最接近的timestamp作为起点
+            closest_idx = min(range(len(sorted_timestamps)), key=lambda i: abs(sorted_timestamps[i] - timestamp))
+
+            # 对数功率谱相关性计算
+            def log_powerSpectrum_corr(audio1, audio2):
+                window = np.hanning(len(audio1))
+                fft1 = np.fft.rfft(audio1 * window)
+                fft2 = np.fft.rfft(audio2 * window)
+                psd1 = np.abs(fft1) ** 2
+                psd2 = np.abs(fft2) ** 2
+                log1 = 10 * np.log10(psd1 + 1e-8)
+                log2 = 10 * np.log10(psd2 + 1e-8)
+                P_xy = np.dot(log1, log2)
+                P_xx = np.dot(log1, log1)
+                P_yy = np.dot(log2, log2)
+                return abs(P_xy) / (np.sqrt(P_xx) * np.sqrt(P_yy) + 1e-8)
+
+            # 用对数功率谱匹配找最佳帧：前后各找2帧
+            best_corr = -1
+            best_ref_idx = closest_idx
+
+            for offset in range(-2, 3):  # T-2, T-1, T, T+1, T+2
+                test_idx = closest_idx + offset
+                if test_idx < 0 or test_idx >= len(sorted_timestamps):
+                    continue
+                test_ts = sorted_timestamps[test_idx]
+                test_ref = np.frombuffer(self.aec_audio_cache[test_ts], dtype=np.int16).astype(np.float32)
+                test_ref_rms = np.sqrt(np.mean(test_ref ** 2))
+                if test_ref_rms < 50:
+                    continue
+
+                # 对数功率谱相关性
+                corr = log_powerSpectrum_corr(mic_audio, test_ref)
+
+                if corr > best_corr:
+                    best_corr = corr
+                    best_ref_idx = test_idx
+
+            best_ts = sorted_timestamps[best_ref_idx]
+
+            best_ref = np.frombuffer(self.aec_audio_cache[best_ts], dtype=np.int16).astype(np.float32)
+            ref_rms = np.sqrt(np.mean(best_ref ** 2))
+
+            if ref_rms < 50:
+                return pcm_frame
+
+            # 对齐参考信号（直接截取相同长度）
+            aligned_ref = best_ref[:n]
+            if len(aligned_ref) < n:
+                aligned_ref = np.pad(aligned_ref, (0, n - len(aligned_ref)))
+
+            # ========== 频域 AEC 处理（谱减法） ==========
+            # 时域信号经过声学路径后相位失真，导致时域相关性低且P_xy正负不定
+            # 频域幅度谱不受相位影响，对数功率谱相关性稳定在0.97+
+            # 公式：result_mag = max(|mic_fft| - |ref_fft| * scale * coef, 0)
+
+            window = np.hanning(n)
+            mic_fft = np.fft.rfft(mic_audio * window)
+            ref_fft = np.fft.rfft(aligned_ref * window)
+
+            mic_mag = np.abs(mic_fft)
+            ref_mag = np.abs(ref_fft)
+            mic_phase = np.angle(mic_fft)
+
+            # 频域计算回声比例 scale
+            scale = np.sum(mic_mag * ref_mag) / (np.dot(ref_mag, ref_mag) + 1e-8)
+
+            # 使用对数功率谱相关性作为coh
+            coh = best_corr
+            ref_power = ref_rms
+
+            # 自适应系数：根据scale和coh动态调整
+            # scale大（回声强）-> coef大；coh高（匹配准）-> coef大
+            raw_coef = 1.0 + scale * 3 + (coh - 0.97) * 30
+            coef = max(0.5, min(3.0, raw_coef))
+
+            # 谱减法（过减 + 半波整流）
+            if ref_power < 50:
+                output = mic_audio
+                coef = 0
+                gain = 0
             else:
-                self.asr_audio_queue.put(audio_data)
+                echo_mag = ref_mag * scale * coef
+                # 过减：使用较大系数抑制回声
+                over_subtract = 1.5
+                result_mag = np.maximum(mic_mag - echo_mag * over_subtract, mic_mag * 0.1)
+
+                # 保留相位重建信号
+                result_fft = result_mag * np.exp(1j * mic_phase)
+                output = np.fft.irfft(result_fft, n)
+                gain = scale * coef
+
+                # 高置信度是纯回声时，再压一下确保VAD检测不到
+                if coh >= 0.97 and ref_power > 500:
+                    output = output * 0.3
+
+            # 后处理：限幅
+            output = np.clip(output, -32768, 32767)
+
+            # 转换为bytes
+            result = output.astype(np.int16).tobytes()
+
+            return result
+
+        except Exception as e:
+            self.logger.bind(tag=TAG).warning(f"[AEC] 处理失败: {e}")
+            return pcm_frame
+
+    def _decode_opus_packet(self, opus_packet: bytes) -> bytes:
+        """
+        解码Opus数据包为PCM数据
+
+        Args:
+            opus_packet: Opus编码的音频数据
+
+        Returns:
+            bytes: 解码后的PCM数据，失败返回None
+        """
+        try:
+            if not opus_packet or len(opus_packet) == 0:
+                return None
+
+            self._init_connection_state(self)
+            pcm_frame = self._connection_opus_decoder.decode(opus_packet, 960)
+            return pcm_frame
+        except Exception as e:
+            self.logger.bind(tag=TAG).debug(f"Opus解码失败: {e}")
+            return None
+
+    def _init_connection_state(self, conn):
+        """为连接初始化独立的Opus解码器"""
+        if not hasattr(conn, "_connection_opus_decoder"):
+            conn._connection_opus_decoder = opuslib_next.Decoder(16000, 1)
 
     async def handle_restart(self, message):
         """处理服务器重启请求"""
@@ -1403,6 +1545,13 @@ class ConnectionHandler:
             ):
                 self.vad.release_conn_resources(self)
 
+            # 清理opus解码器
+            if hasattr(self, "_connection_opus_decoder"):
+                try:
+                    delattr(self, "_connection_opus_decoder")
+                except Exception:
+                    pass
+
             # 清理音频缓冲区
             if hasattr(self, "audio_buffer"):
                 self.audio_buffer.clear()
@@ -1416,6 +1565,20 @@ class ConnectionHandler:
                     pass
                 self.timeout_task = None
 
+            # 取消AEC缓存清理任务
+            if hasattr(self, "_aec_cache_cleanup_task") and self._aec_cache_cleanup_task and not self._aec_cache_cleanup_task.done():
+                self._aec_cache_cleanup_task.cancel()
+                try:
+                    await self._aec_cache_cleanup_task
+                except asyncio.CancelledError:
+                    pass
+                self._aec_cache_cleanup_task = None
+
+            # 清理AEC缓存
+            if hasattr(self, "aec_audio_cache"):
+                self.aec_audio_cache.clear()
+                self.aec_audio_cache_time.clear()
+
             # 清理工具处理器资源
             if hasattr(self, "func_handler") and self.func_handler:
                 try:
@@ -1579,6 +1742,26 @@ class ConnectionHandler:
         finally:
             self.logger.bind(tag=TAG).info("超时检查任务已退出")
 
+    async def _check_aec_cache_expiry(self):
+        """定期清理过期的AEC缓存"""
+        try:
+            while not self.stop_event.is_set():
+                if hasattr(self, "aec_audio_cache") and self.aec_audio_cache:
+                    current_time = time.time()
+                    expired_keys = [
+                        ts for ts, cache_time in list(self.aec_audio_cache_time.items())
+                        if current_time - cache_time > 120  # 2分钟过期
+                    ]
+                    for ts in expired_keys:
+                        self.aec_audio_cache.pop(ts, None)
+                        self.aec_audio_cache_time.pop(ts, None)
+                    if expired_keys:
+                        self.logger.bind(tag=TAG).debug(f"[AEC] 清理过期缓存 {len(expired_keys)} 条")
+                # 每30秒检查一次
+                await asyncio.sleep(30)
+        except Exception as e:
+            self.logger.bind(tag=TAG).error(f"AEC缓存清理任务出错: {e}")
+
     @staticmethod
     def _extract_direct_answer_response(arguments_str):
         """从 direct_answer 的参数中提取 response 值。

main/xiaozhi-server/core/handle/helloHandle.py

@@ -56,6 +56,9 @@ async def handleHelloMessage(conn: "ConnectionHandler", msg_json):
             conn.mcp_client = MCPClient()
             # 发送初始化
             asyncio.create_task(send_mcp_initialize_message(conn))
+        if features.get("aec"):
+            conn.logger.bind(tag=TAG).debug("客户端启用了服务端AEC")
+            conn.client_aec = True
 
     await conn.websocket.send(json.dumps(conn.welcome_msg))
 

main/xiaozhi-server/core/handle/receiveAudioHandle.py

@@ -14,9 +14,9 @@ from core.handle.sendAudioHandle import send_stt_message, SentenceType
 TAG = __name__
 
 
-async def handleAudioMessage(conn: "ConnectionHandler", audio):
+async def handleAudioMessage(conn: "ConnectionHandler", pcm_frame):
     # 当前片段是否有人说话
-    have_voice = conn.vad.is_vad(conn, audio)
+    have_voice = conn.vad.is_vad(conn, pcm_frame)
     # 如果设备刚刚被唤醒，短暂忽略VAD检测
     if hasattr(conn, "just_woken_up") and conn.just_woken_up:
         have_voice = False
@@ -24,10 +24,14 @@ async def handleAudioMessage(conn: "ConnectionHandler", audio):
         if not hasattr(conn, "vad_resume_task") or conn.vad_resume_task.done():
             conn.vad_resume_task = asyncio.create_task(resume_vad_detection(conn))
         return
+    # 服务端AEC功能需要实时触发打断
+    if conn.client_aec and have_voice:
+        if conn.client_is_speaking and conn.client_listen_mode != "manual":
+            await handleAbortMessage(conn)
     # 设备长时间空闲检测，用于say goodbye
     await no_voice_close_connect(conn, have_voice)
     # 接收音频
-    await conn.asr.receive_audio(conn, audio, have_voice)
+    await conn.asr.receive_audio(conn, pcm_frame, have_voice)
 
 
 async def resume_vad_detection(conn: "ConnectionHandler"):

main/xiaozhi-server/core/handle/reportHandle.py

@@ -50,33 +50,27 @@ async def report(conn: "ConnectionHandler", type, text, opus_data, report_time):
         conn.logger.bind(tag=TAG).error(f"聊天记录上报失败: {e}")
 
 
-def opus_to_wav(conn: "ConnectionHandler", opus_data):
-    """将Opus数据转换为WAV格式的字节流
+def opus_to_wav(conn: "ConnectionHandler", pcm_data):
+    """将PCM数据转换为WAV格式的字节流
 
     Args:
         output_dir: 输出目录（保留参数以保持接口兼容）
-        opus_data: opus音频数据
+        pcm_data: PCM音频数据（可能是列表或bytes）
 
     Returns:
         bytes: WAV格式的音频数据
     """
-    decoder = None
     try:
-        decoder = opuslib_next.Decoder(16000, 1)  # 16kHz, 单声道
-        pcm_data = []
+        # 处理可能是列表或bytes的PCM数据
+        if isinstance(pcm_data, list):
+            pcm_data_bytes = b"".join(pcm_data)
+        else:
+            pcm_data_bytes = 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:
-                conn.logger.bind(tag=TAG).error(f"Opus解码错误: {e}", exc_info=True)
-
-        if not pcm_data:
+        if not pcm_data_bytes:
             raise ValueError("没有有效的PCM数据")
 
         # 创建WAV文件头
-        pcm_data_bytes = b"".join(pcm_data)
         num_samples = len(pcm_data_bytes) // 2  # 16-bit samples
 
         # WAV文件头
@@ -97,12 +91,9 @@ def opus_to_wav(conn: "ConnectionHandler", opus_data):
 
         # 返回完整的WAV数据
         return bytes(wav_header) + pcm_data_bytes
-    finally:
-        if decoder is not None:
-            try:
-                del decoder
-            except Exception as e:
-                conn.logger.bind(tag=TAG).debug(f"释放decoder资源时出错: {e}")
+    except Exception as e:
+        conn.logger.bind(tag=TAG).error(f"PCM转WAV失败: {e}", exc_info=True)
+        raise
 
 
 def enqueue_tts_report(conn: "ConnectionHandler", text, opus_data):

main/xiaozhi-server/core/handle/sendAudioHandle.py

@@ -1,6 +1,7 @@
 import json
 import time
 import asyncio
+import opuslib_next
 from typing import TYPE_CHECKING
 
 if TYPE_CHECKING:
@@ -81,13 +82,24 @@ async def _send_to_mqtt_gateway(
     conn: "ConnectionHandler", opus_packet, timestamp, sequence
 ):
     """
-    发送带16字节头部的opus数据包给mqtt_gateway
+    发送带16字节头部的opus数据包给mqtt_gateway，同时缓存音频用于AEC处理
     Args:
         conn: 连接对象
         opus_packet: opus数据包
         timestamp: 时间戳
         sequence: 序列号
     """
+    # 如果启用了服务端AEC，缓存PCM数据用于后续AEC处理
+    if conn.client_aec and timestamp > 0:
+        if not hasattr(conn, "aec_audio_cache"):
+            conn.aec_audio_cache = {}
+            conn.aec_audio_cache_time = {}
+            conn._send_opus_decoder = opuslib_next.Decoder(16000, 1)
+        # 解码opus为PCM后缓存
+        pcm_data = conn._send_opus_decoder.decode(bytes(opus_packet), 960)
+        conn.aec_audio_cache[timestamp] = bytes(pcm_data)
+        conn.aec_audio_cache_time[timestamp] = time.time()
+
     # 为opus数据包添加16字节头部
     header = bytearray(16)
     header[0] = 1  # type

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

@@ -213,7 +213,7 @@ class ASRProvider(ASRProviderBase):
             return None
 
     async def speech_to_text(
-        self, opus_data: List[bytes], session_id: str, audio_format="opus", artifacts=None
+        self, opus_data: List[bytes], session_id: str, artifacts=None
     ) -> Tuple[Optional[str], Optional[str]]:
         """将语音数据转换为文本"""
         if self._is_token_expired():

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

@@ -7,7 +7,6 @@ import hashlib
 import asyncio
 import requests
 import websockets
-import opuslib_next
 from urllib import parse
 from datetime import datetime
 from config.logger import setup_logging
@@ -74,7 +73,6 @@ class ASRProvider(ASRProviderBase):
         self.interface_type = InterfaceType.STREAM
         self.config = config
         self.text = ""
-        self.decoder = opuslib_next.Decoder(16000, 1)
         self.asr_ws = None
         self.forward_task = None
         self.is_processing = False
@@ -129,9 +127,9 @@ class ASRProvider(ASRProviderBase):
     async def open_audio_channels(self, conn):
         await super().open_audio_channels(conn)
 
-    async def receive_audio(self, conn, audio, audio_have_voice):
+    async def receive_audio(self, conn, pcm_frame, audio_have_voice):
         # 先调用父类方法处理基础逻辑
-        await super().receive_audio(conn, audio, audio_have_voice)
+        await super().receive_audio(conn, pcm_frame, audio_have_voice)
 
         # 只在有声音且没有连接时建立连接（排除正在停止的情况）
         if audio_have_voice and not self.is_processing and not self.asr_ws:
@@ -144,7 +142,6 @@ class ASRProvider(ASRProviderBase):
 
         if self.asr_ws and self.is_processing and self.server_ready:
             try:
-                pcm_frame = self.decoder.decode(audio, 960)
                 await self.asr_ws.send(pcm_frame)
             except Exception as e:
                 logger.bind(tag=TAG).warning(f"发送音频失败: {str(e)}")
@@ -232,10 +229,9 @@ class ASRProvider(ASRProviderBase):
 
                         # 发送缓存音频
                         if conn.asr_audio:
-                            for cached_audio in conn.asr_audio[-10:]:
+                            for cached_pcm in conn.asr_audio[-10:]:
                                 try:
-                                    pcm_frame = self.decoder.decode(cached_audio, 960)
-                                    await self.asr_ws.send(pcm_frame)
+                                    await self.asr_ws.send(cached_pcm)
                                 except Exception as e:
                                     logger.bind(tag=TAG).warning(f"发送缓存音频失败: {e}")
                                     break
@@ -328,7 +324,7 @@ class ASRProvider(ASRProviderBase):
 
         logger.bind(tag=TAG).debug("ASR会话清理完成")
 
-    async def speech_to_text(self, opus_data, session_id, audio_format, artifacts=None):
+    async def speech_to_text(self, opus_data, session_id, artifacts=None):
         """获取识别结果"""
         result = self.text
         self.text = ""
@@ -337,10 +333,3 @@ class ASRProvider(ASRProviderBase):
     async def close(self):
         """关闭资源"""
         await self._cleanup()
-        if hasattr(self, 'decoder') and self.decoder is not None:
-            try:
-                del self.decoder
-                self.decoder = None
-                logger.bind(tag=TAG).debug("Aliyun decoder resources released")
-            except Exception as e:
-                logger.bind(tag=TAG).debug(f"释放Aliyun decoder资源时出错: {e}")

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

@@ -2,7 +2,6 @@ import json
 import uuid
 import asyncio
 import websockets
-import opuslib_next
 from typing import List, TYPE_CHECKING
 
 if TYPE_CHECKING:
@@ -22,7 +21,6 @@ class ASRProvider(ASRProviderBase):
         self.interface_type = InterfaceType.STREAM
         self.config = config
         self.text = ""
-        self.decoder = opuslib_next.Decoder(16000, 1)
         self.asr_ws = None
         self.forward_task = None
         self.is_processing = False
@@ -55,9 +53,9 @@ class ASRProvider(ASRProviderBase):
     async def open_audio_channels(self, conn):
         await super().open_audio_channels(conn)
 
-    async def receive_audio(self, conn, audio, audio_have_voice):
+    async def receive_audio(self, conn, pcm_frame, audio_have_voice):
         # 先调用父类方法处理基础逻辑
-        await super().receive_audio(conn, audio, audio_have_voice)
+        await super().receive_audio(conn, pcm_frame, audio_have_voice)
 
         # 只在有声音且没有连接时建立连接
         if audio_have_voice and not self.is_processing and not self.asr_ws:
@@ -71,8 +69,6 @@ class ASRProvider(ASRProviderBase):
         # 发送音频数据
         if self.asr_ws and self.is_processing and self.server_ready:
             try:
-                pcm_frame = self.decoder.decode(audio, 960)
-                # 直接发送PCM音频数据(二进制)
                 await self.asr_ws.send(pcm_frame)
             except Exception as e:
                 logger.bind(tag=TAG).warning(f"发送音频失败: {str(e)}")
@@ -179,10 +175,9 @@ class ASRProvider(ASRProviderBase):
 
                         # 发送缓存音频
                         if conn.asr_audio:
-                            for cached_audio in conn.asr_audio[-10:]:
+                            for cached_pcm in conn.asr_audio[-10:]:
                                 try:
-                                    pcm_frame = self.decoder.decode(cached_audio, 960)
-                                    await self.asr_ws.send(pcm_frame)
+                                    await self.asr_ws.send(cached_pcm)
                                 except Exception as e:
                                     logger.bind(tag=TAG).warning(f"发送缓存音频失败: {e}")
                                     break
@@ -312,7 +307,7 @@ class ASRProvider(ASRProviderBase):
 
         logger.bind(tag=TAG).debug("ASR会话清理完成")
 
-    async def speech_to_text(self, opus_data, session_id, audio_format, artifacts=None):
+    async def speech_to_text(self, opus_data, session_id, artifacts=None):
         """获取识别结果"""
         result = self.text
         self.text = ""
@@ -320,11 +315,4 @@ class ASRProvider(ASRProviderBase):
 
     async def close(self):
         """关闭资源"""
-        await self._cleanup()
-        if hasattr(self, 'decoder') and self.decoder is not None:
-            try:
-                del self.decoder
-                self.decoder = None
-                logger.bind(tag=TAG).debug("Aliyun BL decoder resources released")
-            except Exception as e:
-                logger.bind(tag=TAG).debug(f"释放Aliyun BL decoder资源时出错: {e}")
+        await self._cleanup()

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

@@ -30,7 +30,7 @@ class ASRProvider(ASRProviderBase):
         os.makedirs(self.output_dir, exist_ok=True)
 
     async def speech_to_text(
-        self, opus_data: List[bytes], session_id: str, audio_format="opus", artifacts=None
+        self, opus_data: List[bytes], session_id: str, artifacts=None
     ) -> Tuple[Optional[str], Optional[str]]:
         """将语音数据转换为文本"""
         if not opus_data:

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

@@ -10,7 +10,6 @@ import asyncio
 import tempfile
 import traceback
 import threading
-import opuslib_next
 
 from abc import ABC, abstractmethod
 from config.logger import setup_logging
@@ -59,13 +58,13 @@ class ASRProviderBase(ABC):
                 continue
 
     # 接收音频
-    async def receive_audio(self, conn: "ConnectionHandler", audio, audio_have_voice):
+    async def receive_audio(self, conn: "ConnectionHandler", pcm_frame, audio_have_voice):
         if conn.client_listen_mode == "manual":
             # 手动模式：缓存音频用于ASR识别
-            conn.asr_audio.append(audio)
+            conn.asr_audio.append(pcm_frame)
         else:
             # 自动/实时模式：使用VAD检测
-            conn.asr_audio.append(audio)
+            conn.asr_audio.append(pcm_frame)
 
             # 如果没有语音，且之前也没有声音，缓存部分音频
             if not audio_have_voice and not conn.client_have_voice:
@@ -74,24 +73,21 @@ class ASRProviderBase(ABC):
 
             # 自动模式下通过VAD检测到语音停止时触发识别
             if conn.asr.interface_type != InterfaceType.STREAM and conn.client_voice_stop:
-                asr_audio_task = conn.asr_audio.copy()
+                # 直接使用asr_audio中的PCM数据
+                pcm_bytes = b"".join(conn.asr_audio)
+                # 检查是否有足够的音频数据（每帧1920字节，15帧约28800字节）
+                if len(pcm_bytes) > 1920 * 15:
+                    await self.handle_voice_stop(conn, [pcm_bytes])
                 conn.reset_audio_states()
 
-                if len(asr_audio_task) > 15:
-                    await self.handle_voice_stop(conn, asr_audio_task)
-
     # 处理语音停止
     async def handle_voice_stop(self, conn: "ConnectionHandler", asr_audio_task: List[bytes]):
         """并行处理ASR和声纹识别"""
         try:
             total_start_time = time.monotonic()
 
-            # 准备音频数据
-            if conn.audio_format == "pcm":
-                pcm_data = asr_audio_task
-            else:
-                pcm_data = self.decode_opus(asr_audio_task)
-
+            # 数据已经是PCM直接使用
+            pcm_data = asr_audio_task
             combined_pcm_data = b"".join(pcm_data)
 
             # 预先准备WAV数据
@@ -101,7 +97,7 @@ class ASRProviderBase(ABC):
 
             # 定义ASR任务
             asr_task = self.speech_to_text_wrapper(
-                asr_audio_task, conn.session_id, conn.audio_format
+                asr_audio_task, conn.session_id
             )
 
             if conn.voiceprint_provider and wav_data:
@@ -270,15 +266,11 @@ class ASRProviderBase(ABC):
         return file_path
 
     async def speech_to_text_wrapper(
-        self, opus_data: List[bytes], session_id: str, audio_format="opus"
+        self, pcm_data: List[bytes], session_id: str
     ) -> Tuple[Optional[str], Optional[str]]:
         file_path = None
         temp_path = None
         try:
-            if audio_format == "pcm":
-                pcm_data = opus_data
-            else:
-                pcm_data = self.decode_opus(opus_data)
             combined_pcm_data = b"".join(pcm_data)
 
             free_space = shutil.disk_usage(self.output_dir).free
@@ -304,7 +296,7 @@ class ASRProviderBase(ABC):
                 )
 
             text, _ = await self.speech_to_text(
-                opus_data, session_id, audio_format, artifacts
+                pcm_data, session_id, artifacts
             )
             return text, file_path
         except OSError as e:
@@ -332,50 +324,13 @@ class ASRProviderBase(ABC):
         self,
         opus_data: List[bytes],
         session_id: str,
-        audio_format="opus",
         artifacts: Optional[AudioArtifacts] = None,
     ) -> Tuple[Optional[str], Optional[str]]:
         """将语音数据转换为文本
 
         :param opus_data: 输入的Opus音频数据
         :param session_id: 会话ID
-        :param audio_format: 音频格式，默认"opus"
         :param artifacts: 音频工件，包含PCM数据、文件路径等
         :return: 识别结果文本和文件路径（如果有）
         """
         pass
-
-    @staticmethod
-    def decode_opus(opus_data: List[bytes]) -> List[bytes]:
-        """将Opus音频数据解码为PCM数据"""
-        decoder = None
-        try:
-            decoder = opuslib_next.Decoder(16000, 1)
-            pcm_data = []
-            buffer_size = 960  # 每次处理960个采样点 (60ms at 16kHz)
-
-            for i, opus_packet in enumerate(opus_data):
-                try:
-                    if not opus_packet or len(opus_packet) == 0:
-                        continue
-
-                    pcm_frame = decoder.decode(opus_packet, buffer_size)
-                    if pcm_frame and len(pcm_frame) > 0:
-                        pcm_data.append(pcm_frame)
-
-                except opuslib_next.OpusError as e:
-                    logger.bind(tag=TAG).warning(f"Opus解码错误，跳过数据包 {i}: {e}")
-                except Exception as e:
-                    logger.bind(tag=TAG).error(f"音频处理错误，数据包 {i}: {e}")
-
-            return pcm_data
-
-        except Exception as e:
-            logger.bind(tag=TAG).error(f"音频解码过程发生错误: {e}")
-            return []
-        finally:
-            if decoder is not None:
-                try:
-                    del decoder
-                except Exception as e:
-                    logger.bind(tag=TAG).debug(f"释放decoder资源时出错: {e}")

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

@@ -232,7 +232,7 @@ class ASRProvider(ASRProviderBase):
             yield data[offset:data_len], True
 
     async def speech_to_text(
-        self, opus_data: List[bytes], session_id: str, audio_format="opus", artifacts=None
+        self, opus_data: List[bytes], session_id: str, artifacts=None
     ) -> Tuple[Optional[str], Optional[str]]:
         """将语音数据转换为文本"""
 

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

@@ -3,7 +3,6 @@ import gzip
 import uuid
 import asyncio
 import websockets
-import opuslib_next
 from core.providers.asr.base import ASRProviderBase
 from config.logger import setup_logging
 from core.providers.asr.dto.dto import InterfaceType
@@ -22,7 +21,6 @@ class ASRProvider(ASRProviderBase):
         self.interface_type = InterfaceType.STREAM
         self.config = config
         self.text = ""
-        self.decoder = opuslib_next.Decoder(16000, 1)
         self.asr_ws = None
         self.forward_task = None
         self.is_processing = False  # 添加处理状态标志
@@ -68,10 +66,10 @@ class ASRProvider(ASRProviderBase):
     async def open_audio_channels(self, conn):
         await super().open_audio_channels(conn)
 
-    async def receive_audio(self, conn: "ConnectionHandler", audio, audio_have_voice):
+    async def receive_audio(self, conn: "ConnectionHandler", pcm_frame, audio_have_voice):
         # 先调用父类方法处理基础逻辑
-        await super().receive_audio(conn, audio, audio_have_voice)
-        
+        await super().receive_audio(conn, pcm_frame, audio_have_voice)
+
         # 如果本次有声音，且之前没有建立连接
         if audio_have_voice and self.asr_ws is None and not self.is_processing:
             try:
@@ -123,10 +121,9 @@ class ASRProvider(ASRProviderBase):
 
                 # 发送缓存的音频数据
                 if conn.asr_audio and len(conn.asr_audio) > 0:
-                    for cached_audio in conn.asr_audio[-10:]:
+                    for cached_pcm in conn.asr_audio[-10:]:
                         try:
-                            pcm_frame = self.decoder.decode(cached_audio, 960)
-                            payload = gzip.compress(pcm_frame)
+                            payload = gzip.compress(cached_pcm)
                             audio_request = bytearray(
                                 self.generate_audio_default_header()
                             )
@@ -151,7 +148,6 @@ class ASRProvider(ASRProviderBase):
         # 发送当前音频数据
         if self.asr_ws and self.is_processing and not self._is_stopping:
             try:
-                pcm_frame = self.decoder.decode(audio, 960)
                 payload = gzip.compress(pcm_frame)
                 audio_request = bytearray(self.generate_audio_default_header())
                 audio_request.extend(len(payload).to_bytes(4, "big"))
@@ -414,7 +410,7 @@ class ASRProvider(ASRProviderBase):
             logger.bind(tag=TAG).error(f"原始响应数据: {res.hex()}")
             raise
 
-    async def speech_to_text(self, opus_data, session_id, audio_format, artifacts=None):
+    async def speech_to_text(self, opus_data, session_id, artifacts=None):
         result = self.text
         self.text = ""  # 清空text
         return result, None
@@ -433,11 +429,3 @@ class ASRProvider(ASRProviderBase):
             self.forward_task = None
         self.is_processing = False
 
-        # 显式释放decoder资源
-        if hasattr(self, "decoder") and self.decoder is not None:
-            try:
-                del self.decoder
-                self.decoder = None
-                logger.bind(tag=TAG).debug("Doubao decoder resources released")
-            except Exception as e:
-                logger.bind(tag=TAG).debug(f"释放Doubao decoder资源时出错: {e}")

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

@@ -65,7 +65,7 @@ class ASRProvider(ASRProviderBase):
             )
 
     async def speech_to_text(
-        self, opus_data: List[bytes], session_id: str, audio_format="opus", artifacts=None
+        self, opus_data: List[bytes], session_id: str, artifacts=None
     ) -> Tuple[Optional[str], Optional[str]]:
         """语音转文本主处理逻辑"""
         retry_count = 0

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

@@ -101,7 +101,7 @@ class ASRProvider(ASRProviderBase):
         logger.bind(tag=TAG).debug(f"Sent end message: {end_message}")
 
     async def speech_to_text(
-        self, opus_data: List[bytes], session_id: str, audio_format="opus", artifacts=None
+        self, opus_data: List[bytes], session_id: str, artifacts=None
     ) -> Tuple[Optional[str], Optional[str]]:
         """
         Convert speech data to text using FunASR.

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

@@ -24,7 +24,7 @@ class ASRProvider(ASRProviderBase):
     def requires_file(self) -> bool:
         return True
 
-    async def speech_to_text(self, opus_data: List[bytes], session_id: str, audio_format="opus", artifacts=None) -> Tuple[Optional[str], Optional[str]]:
+    async def speech_to_text(self, opus_data: List[bytes], session_id: str, artifacts=None) -> Tuple[Optional[str], Optional[str]]:
         file_path = None
         try:
             if artifacts is None:

main/xiaozhi-server/core/providers/asr/qwen3_asr_flash.py

@@ -41,7 +41,7 @@ class ASRProvider(ASRProviderBase):
         return True
 
     async def speech_to_text(
-        self, opus_data: List[bytes], session_id: str, audio_format="opus", artifacts=None
+        self, opus_data: List[bytes], session_id: str, artifacts=None
     ) -> Tuple[Optional[str], Optional[str]]:
         """将语音数据转换为文本"""
         temp_file_path = None

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

@@ -124,7 +124,7 @@ class ASRProvider(ASRProviderBase):
         return True
 
     async def speech_to_text(
-        self, opus_data: List[bytes], session_id: str, audio_format="opus", artifacts=None
+        self, opus_data: List[bytes], session_id: str, artifacts=None
     ) -> Tuple[Optional[str], Optional[str]]:
         """语音转文本主处理逻辑"""
         file_path = None

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

@@ -32,7 +32,7 @@ class ASRProvider(ASRProviderBase):
         os.makedirs(self.output_dir, exist_ok=True)
 
     async def speech_to_text(
-        self, opus_data: List[bytes], session_id: str, audio_format="opus", artifacts=None
+        self, opus_data: List[bytes], session_id: str, artifacts=None
     ) -> Tuple[Optional[str], Optional[str]]:
         """将语音数据转换为文本"""
         if not opus_data:

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

@@ -44,7 +44,7 @@ class ASRProvider(ASRProviderBase):
             raise
 
     async def speech_to_text(
-        self, opus_data: List[bytes], session_id: str, audio_format="opus", artifacts=None
+        self, opus_data: List[bytes], session_id: str, artifacts=None
     ) -> Tuple[Optional[str], Optional[str]]:
         """将语音数据转换为文本"""
         try:

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

@@ -4,7 +4,6 @@ import base64
 import hashlib
 import asyncio
 import websockets
-import opuslib_next
 import gc
 from time import mktime
 from datetime import datetime
@@ -33,7 +32,6 @@ class ASRProvider(ASRProviderBase):
         self.interface_type = InterfaceType.STREAM
         self.config = config
         self.text = ""
-        self.decoder = opuslib_next.Decoder(16000, 1)
         self.asr_ws = None
         self.forward_task = None
         self.is_processing = False
@@ -100,9 +98,9 @@ class ASRProvider(ASRProviderBase):
     async def open_audio_channels(self, conn: "ConnectionHandler"):
         await super().open_audio_channels(conn)
 
-    async def receive_audio(self, conn: "ConnectionHandler", audio, audio_have_voice):
+    async def receive_audio(self, conn: "ConnectionHandler", pcm_frame, audio_have_voice):
         # 先调用父类方法处理基础逻辑
-        await super().receive_audio(conn, audio, audio_have_voice)
+        await super().receive_audio(conn, pcm_frame, audio_have_voice)
 
         # 如果本次有声音，且之前没有建立连接
         if audio_have_voice and self.asr_ws is None and not self.is_processing:
@@ -116,7 +114,6 @@ class ASRProvider(ASRProviderBase):
         # 发送当前音频数据
         if self.asr_ws and self.is_processing and self.server_ready:
             try:
-                pcm_frame = self.decoder.decode(audio, 960)
                 await self._send_audio_frame(pcm_frame, STATUS_CONTINUE_FRAME)
             except Exception as e:
                 logger.bind(tag=TAG).warning(f"发送音频数据时发生错误: {e}")
@@ -148,19 +145,15 @@ class ASRProvider(ASRProviderBase):
 
             # 发送首帧音频
             if conn.asr_audio and len(conn.asr_audio) > 0:
-                first_audio = conn.asr_audio[-1] if conn.asr_audio else b""
-                pcm_frame = (
-                    self.decoder.decode(first_audio, 960) if first_audio else b""
-                )
-                await self._send_audio_frame(pcm_frame, STATUS_FIRST_FRAME)
+                first_pcm = conn.asr_audio[-1] if conn.asr_audio else b""
+                await self._send_audio_frame(first_pcm, STATUS_FIRST_FRAME)
                 self.server_ready = True
                 logger.bind(tag=TAG).info("已发送首帧，开始识别")
 
                 # 发送缓存的音频数据
-                for cached_audio in conn.asr_audio[-10:]:
+                for cached_pcm in conn.asr_audio[-10:]:
                     try:
-                        pcm_frame = self.decoder.decode(cached_audio, 960)
-                        await self._send_audio_frame(pcm_frame, STATUS_CONTINUE_FRAME)
+                        await self._send_audio_frame(cached_pcm, STATUS_CONTINUE_FRAME)
                     except Exception as e:
                         logger.bind(tag=TAG).info(f"发送缓存音频数据时发生错误: {e}")
                         break
@@ -322,7 +315,7 @@ class ASRProvider(ASRProviderBase):
 
         logger.bind(tag=TAG).debug("ASR会话清理完成")
 
-    async def speech_to_text(self, opus_data, session_id, audio_format, artifacts=None):
+    async def speech_to_text(self, opus_data, session_id, artifacts=None):
         """获取识别结果"""
         result = self.text
         self.text = ""
@@ -342,12 +335,3 @@ class ASRProvider(ASRProviderBase):
             self.forward_task = None
         self.is_processing = False
 
-        # 显式释放decoder资源
-        if hasattr(self, "decoder") and self.decoder is not None:
-            try:
-                del self.decoder
-                self.decoder = None
-                logger.bind(tag=TAG).debug("Xunfei decoder resources released")
-            except Exception as e:
-                logger.bind(tag=TAG).debug(f"释放Xunfei decoder资源时出错: {e}")
-

main/xiaozhi-server/core/providers/vad/silero.py

@@ -1,7 +1,6 @@
 import time
 import os
 import numpy as np
-import opuslib_next
 import onnxruntime
 from config.logger import setup_logging
 from core.providers.vad.base import VADProviderBase
@@ -39,8 +38,6 @@ class VADProvider(VADProviderBase):
 
     def _init_connection_state(self, conn):
         """为连接初始化独立的 VAD 状态"""
-        if not hasattr(conn, "_vad_opus_decoder"):
-            conn._vad_opus_decoder = opuslib_next.Decoder(16000, 1)
         if not hasattr(conn, "_vad_state"):
             conn._vad_state = np.zeros((2, 1, 128), dtype=np.float32)
         if not hasattr(conn, "_vad_context"):
@@ -48,14 +45,14 @@ class VADProvider(VADProviderBase):
 
     def release_conn_resources(self, conn):
         """释放连接的 VAD 资源（连接关闭时调用）"""
-        for attr in ("_vad_opus_decoder", "_vad_state", "_vad_context"):
+        for attr in ("_vad_state", "_vad_context"):
             if hasattr(conn, attr):
                 try:
                     delattr(conn, attr)
                 except Exception:
                     pass
 
-    def is_vad(self, conn, opus_packet):
+    def is_vad(self, conn, pcm_frame):
         # 手动模式：直接返回True，不进行实时VAD检测，所有音频都缓存
         if conn.client_listen_mode == "manual":
             return True
@@ -63,7 +60,7 @@ class VADProvider(VADProviderBase):
         try:
             self._init_connection_state(conn)
 
-            pcm_frame = conn._vad_opus_decoder.decode(opus_packet, 960)
+            # pcm_frame已经是处理后的PCM数据
             conn.client_audio_buffer.extend(pcm_frame)
 
             client_have_voice = False
@@ -115,7 +112,5 @@ class VADProvider(VADProviderBase):
                     conn.vad_last_voice_time = time.time() * 1000
 
             return client_have_voice
-        except opuslib_next.OpusError as e:
-            logger.bind(tag=TAG).info(f"解码错误: {e}")
         except Exception as e:
             logger.bind(tag=TAG).error(f"Error processing audio packet: {e}")