Merge pull request #3436 from xinnan-tech/perf/hello-handshake-fast

hello 握手首屏响应从 4s 降到 200ms

main/xiaozhi-server/app.py

@@ -10,6 +10,7 @@ from core.http_server import SimpleHttpServer
 from core.websocket_server import WebSocketServer
 from core.utils.util import check_ffmpeg_installed
 from core.utils.gc_manager import get_gc_manager
+from core.handle.helloHandle import preheat_wakeup_audio_cache
 
 TAG = __name__
 logger = setup_logging()
@@ -64,6 +65,9 @@ async def main():
     # 添加 stdin 监控任务
     stdin_task = asyncio.create_task(monitor_stdin())
 
+    # 预热唤醒词音频编码缓存，让首条连接就命中
+    asyncio.create_task(preheat_wakeup_audio_cache())
+
     # 启动全局GC管理器（5分钟清理一次）
     gc_manager = get_gc_manager(interval_seconds=300)
     await gc_manager.start()

main/xiaozhi-server/core/connection.py

@@ -781,7 +781,7 @@ class ConnectionHandler:
                     self.voiceprint_provider = voiceprint_provider
                     self.logger.bind(tag=TAG).info("声纹识别功能已在连接时动态启用")
                 else:
-                    self.logger.bind(tag=TAG).warning("声纹识别功能启用但配置不完整")
+                    self.logger.bind(tag=TAG).debug("声纹识别功能启用但配置不完整")
             else:
                 self.logger.bind(tag=TAG).info("声纹识别功能未启用")
         except Exception as e:

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

@@ -4,6 +4,7 @@ import uuid
 import random
 import asyncio
 from typing import TYPE_CHECKING
+from config.logger import setup_logging
 
 if TYPE_CHECKING:
     from core.connection import ConnectionHandler
@@ -13,9 +14,14 @@ from core.providers.tts.dto.dto import SentenceType
 from core.utils.wakeup_word import WakeupWordsConfig
 from core.handle.sendAudioHandle import sendAudioMessage, send_tts_message
 from core.utils.util import remove_punctuation_and_length, opus_datas_to_wav_bytes
-from core.providers.tools.device_mcp import MCPClient, send_mcp_initialize_message
+from core.providers.tools.device_mcp import (
+    MCPClient,
+    send_mcp_initialize_message,
+    send_mcp_tools_list_request,
+)
 
 TAG = __name__
+_wakeup_logger = setup_logging()
 
 WAKEUP_CONFIG = {
     "refresh_time": 10,
@@ -63,6 +69,8 @@ async def handleHelloMessage(conn: "ConnectionHandler", msg_json):
     # The device waits for the server hello before processing MCP messages.
     if features and features.get("mcp"):
         asyncio.create_task(send_mcp_initialize_message(conn))
+        # 并发预发 tools/list，让帧排在 audio 流之前到达 client
+        asyncio.create_task(send_mcp_tools_list_request(conn))
 
 
 async def checkWakeupWords(conn: "ConnectionHandler", text):
@@ -105,19 +113,19 @@ async def checkWakeupWords(conn: "ConnectionHandler", text):
         }
 
     # 获取音频数据
-    opus_packets = await audio_to_data(response.get("file_path"), use_cache=False)
+    opus_packets = await audio_to_data(response.get("file_path"), use_cache=True)
     # 播放唤醒词回复
     conn.client_abort = False
 
     # 将唤醒词回复视为新会话，生成新的 sentence_id，确保流控器重置
     conn.sentence_id = str(uuid.uuid4().hex)
 
-    conn.logger.bind(tag=TAG).info(f"播放唤醒词回复: {response.get('text')}")
-    await sendAudioMessage(conn, SentenceType.FIRST, opus_packets, response.get("text"))
-    await sendAudioMessage(conn, SentenceType.LAST, [], None)
-
     # 补充对话
     conn.dialogue.put(Message(role="assistant", content=response.get("text")))
+    conn.logger.bind(tag=TAG).info(f"播放唤醒词回复: {response.get('text')}")
+
+    await sendAudioMessage(conn, SentenceType.FIRST, opus_packets, response.get("text"))
+    asyncio.create_task(_finish_wakeup_reply(conn))
 
     # 检查是否需要更新唤醒词回复
     if time.time() - response.get("time", 0) > WAKEUP_CONFIG["refresh_time"]:
@@ -126,6 +134,16 @@ async def checkWakeupWords(conn: "ConnectionHandler", text):
     return True
 
 
+async def _finish_wakeup_reply(conn: "ConnectionHandler"):
+    """等音频队列清空后再发 stop（时序：全部音频帧→stop），期间不阻塞接收循环"""
+    try:
+        await sendAudioMessage(conn, SentenceType.LAST, [], None)
+    except asyncio.CancelledError:
+        raise
+    except Exception as e:
+        conn.logger.bind(tag=TAG).error(f"唤醒词回复收尾失败: {e}")
+
+
 async def wakeupWordsResponse(conn: "ConnectionHandler"):
     if not conn.tts:
         return
@@ -160,7 +178,32 @@ async def wakeupWordsResponse(conn: "ConnectionHandler"):
             f.write(wav_bytes)
         # 更新配置
         wakeup_words_config.update_wakeup_response(voice, file_path, result)
+        # 覆写文件后 mtime 变化使旧缓存失效，此处立刻预编码新文件，
+        # 下次唤醒直接命中缓存，编码成本不占用唤醒响应路径
+        await audio_to_data(file_path, use_cache=True)
     finally:
         # 确保在任何情况下都释放锁
         if _wakeup_response_lock.locked():
             _wakeup_response_lock.release()
+
+
+async def preheat_wakeup_audio_cache():
+    """进程启动时预热唤醒词音频编码缓存
+
+    把 wakeup_words/{voice}.wav 在 Opus 编码结果一次性算好塞进全局 cache，
+    后续每条连接上来就能直接命中，省掉 wav→PCM→Opus 的数百毫秒开销。
+    """
+    # 进程启动时 tts 还未初始化，无法拿到 conn.tts.voice，按常见 fallback 列表预热
+    preheat_voices = ["default", "female-shaonv", "male-qn-jingying"]
+    fallback_path = "config/assets/wakeup_words_short.wav"
+    for voice in preheat_voices:
+        file_path = fallback_path
+        try:
+            response = wakeup_words_config.get_wakeup_response(voice)
+            if response and response.get("file_path"):
+                file_path = response["file_path"]
+            await audio_to_data(file_path, use_cache=True)
+        except Exception as e:
+            _wakeup_logger.bind(tag=TAG).warning(
+                f"预热 voice={voice} 失败: {e}"
+            )

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

@@ -12,10 +12,8 @@ from core.providers.tts.dto.dto import SentenceType
 from core.utils.audioRateController import AudioRateController
 
 TAG = __name__
-# 音频帧时长（毫秒）
-AUDIO_FRAME_DURATION = 60
-# 预缓冲包数量，直接发送以减少延迟
-PRE_BUFFER_COUNT = 5
+AUDIO_FRAME_DURATION = 60  # 音频帧时长（毫秒），与 Opus 编码 60ms/帧对齐
+PRE_BUFFER_COUNT = 2  # 首段发送时不入队的帧数，避免流控延迟拉高首帧到达时间
 
 
 async def sendAudioMessage(conn: "ConnectionHandler", sentenceType, audios, text, sentence_id=None):

main/xiaozhi-server/core/providers/tools/device_mcp/mcp_handler.py

@@ -62,11 +62,6 @@ async def handle_mcp_message(
                 logger.bind(tag=TAG).debug(
                     f"客户端MCP服务器信息: name={name}, version={version}"
                 )
-
-            await asyncio.sleep(1)
-            logger.bind(tag=TAG).debug("初始化完成，开始请求MCP工具列表")
-            await send_mcp_tools_list_request(conn)
-
             return
 
         elif msg_id == 2:  # mcpToolsListID

main/xiaozhi-server/core/providers/tools/server_mcp/mcp_manager.py

@@ -24,7 +24,7 @@ class ServerMCPManager:
         self.config_path = get_project_dir() + "data/.mcp_server_settings.json"
         if not os.path.exists(self.config_path):
             self.config_path = ""
-            logger.bind(tag=TAG).warning(
+            logger.bind(tag=TAG).debug(
                 f"请检查mcp服务配置文件：data/.mcp_server_settings.json"
             )
         self.clients: Dict[str, ServerMCPClient] = {}

main/xiaozhi-server/core/utils/prompt_manager.py

@@ -5,6 +5,7 @@
 
 import os
 import asyncio
+import hashlib
 import threading
 from typing import Dict, Any, TYPE_CHECKING
 
@@ -120,8 +121,9 @@ class PromptManager:
 
         # 使用传入的提示词并缓存（如果有设备ID）
         if device_id:
-            device_cache_key = f"device_prompt:{device_id}"
-            self.cache_manager.set(self.CacheType.DEVICE_PROMPT, device_cache_key, user_prompt)
+            self.cache_manager.set(
+                self.CacheType.DEVICE_PROMPT, device_cache_key, user_prompt
+            )
             self.logger.bind(tag=TAG).debug(f"设备 {device_id} 的提示词已缓存")
 
         self.logger.bind(tag=TAG).info(f"使用快速提示词: {user_prompt[:50]}...")
@@ -253,6 +255,18 @@ class PromptManager:
         if not self.base_prompt_template:
             return user_prompt
 
+        # 缓存键须覆盖所有渲染输入：emoji 开关、设备 prompt 变更否则会命中过期提示词
+        emoji_enabled = kwargs.get("emoji_enabled", True)
+        prompt_hash = hashlib.md5(user_prompt.encode("utf-8")).hexdigest()[:8]
+        enhanced_cache_key = (
+            f"enhanced_prompt:{device_id}:{emoji_enabled}:{prompt_hash}"
+        )
+        cached_prompt = self.cache_manager.get(
+            self.CacheType.DEVICE_PROMPT, enhanced_cache_key
+        )
+        if cached_prompt is not None:
+            return cached_prompt
+
         try:
             # 获取最新的时间信息（不缓存）
             today_date, today_weekday, lunar_date = self._get_current_time_info()
@@ -301,11 +315,13 @@ class PromptManager:
                 *args,
                 **kwargs,
             )
-            device_cache_key = f"device_prompt:{device_id}"
             self.cache_manager.set(
-                self.CacheType.DEVICE_PROMPT, device_cache_key, enhanced_prompt
+                self.CacheType.DEVICE_PROMPT,
+                enhanced_cache_key,
+                enhanced_prompt,
+                ttl=300,
             )
-            self.logger.bind(tag=TAG).info(
+            self.logger.bind(tag=TAG).debug(
                 f"构建增强提示词成功，长度: {len(enhanced_prompt)}"
             )
             return enhanced_prompt

main/xiaozhi-server/core/utils/util.py

@@ -310,8 +310,12 @@ async def audio_to_data(
     from core.utils.cache.manager import cache_manager
     from core.utils.cache.config import CacheType
 
-    # 生成缓存键，包含文件路径和编码类型
-    cache_key = f"{audio_file_path}:{is_opus}"
+    # 缓存键含 mtime：文件被原地覆写时旧缓存自动失效（如唤醒词回复重新 TTS），避免播到过期音频
+    try:
+        mtime = os.path.getmtime(audio_file_path)
+        cache_key = f"{audio_file_path}:{mtime}:{is_opus}"
+    except OSError:
+        cache_key = f"{audio_file_path}:{is_opus}"
 
     # 尝试从缓存获取结果
     if use_cache: