语音优化

services/voice_session_manager.py

@@ -19,6 +19,7 @@ class _StreamTagFilter:
     """
 
     SKIP_TAGS = {"TOOL_CALL", "TOOL_RESPONSE", "THINK", "SOURCE", "REFERENCE"}
+    KNOWN_TAGS = {"ANSWER", "TOOL_CALL", "TOOL_RESPONSE", "THINK", "SOURCE", "REFERENCE", "PREAMBLE", "SUMMARY"}
 
     def __init__(self):
         self.state = "idle"  # idle, answer, skip
@@ -50,8 +51,13 @@ class _StreamTagFilter:
 
                     tag_name = self._pending[bracket_pos + 1:close_pos]
                     self._pending = self._pending[close_pos + 1:]
-                    self.found_any_tag = True
 
+                    if tag_name not in self.KNOWN_TAGS:
+                        if self.state == "answer" or not self.found_any_tag:
+                            output.append(f"[{tag_name}]")
+                        continue
+
+                    self.found_any_tag = True
                     if tag_name == "ANSWER":
                         self.state = "answer"
                     else:
@@ -70,6 +76,9 @@ class _StreamTagFilter:
                     tag_name = self._pending[bracket_pos + 1:close_pos]
                     self._pending = self._pending[close_pos + 1:]
 
+                    if tag_name not in self.KNOWN_TAGS:
+                        continue
+
                     if tag_name == "ANSWER":
                         self.state = "answer"
                     else:
@@ -115,6 +124,11 @@ class VoiceSession:
         self._current_asr_text = ""
         # When True, discard TTS audio from SERVER_ACK (comfort speech period)
         self._is_sending_chat_tts_text = False
+        # Set to True when event 350 fires for chat_tts_text, indicating the TTS segment is done
+        # and next TTS send must use start=True to begin a new session
+        self._tts_segment_done = False
+        # Signaled when event 359 fires (TTS fully completed), used to wait before starting new TTS
+        self._tts_complete_event: asyncio.Event = asyncio.Event()
         self._receive_task: Optional[asyncio.Task] = None
         self._agent_task: Optional[asyncio.Task] = None
 
@@ -252,10 +266,14 @@ class VoiceSession:
                     self._is_sending_chat_tts_text = False
                     logger.info(f"[Voice] Comfort/RAG TTS done, resuming audio forwarding")
 
+                # Mark TTS segment as done so next send uses start=True
+                if tts_type == "chat_tts_text":
+                    self._tts_segment_done = True
+
             elif event == 359:
                 # TTS fully completed (all segments done)
                 logger.info(f"[Voice] TTS fully completed")
-                # await self._emit_status("idle")
+                self._tts_complete_event.set()
 
             elif event in (152, 153):
                 logger.info(f"[Voice] Session finished event: {event}")
@@ -331,26 +349,59 @@ class VoiceSession:
                     if sentence:
                         sentence = self._clean_markdown(sentence)
                     if sentence:
+                        # If previous TTS segment completed (e.g. gap during tool call),
+                        # close old session, wait for TTS delivery to finish, then restart
+                        if tts_started and self._tts_segment_done:
+                            logger.info(f"[Voice] TTS segment done, closing session and waiting for delivery")
+                            await self.realtime_client.chat_tts_text(content="", start=False, end=True)
+                            self._tts_complete_event.clear()
+                            try:
+                                await asyncio.wait_for(self._tts_complete_event.wait(), timeout=10)
+                            except asyncio.TimeoutError:
+                                logger.warning(f"[Voice] Timeout waiting for TTS complete, proceeding anyway")
+                            tts_started = False
+                            self._tts_segment_done = False
+                            logger.info(f"[Voice] TTS delivery done, starting new session")
+
                         logger.info(f"[Voice] Sending TTS sentence: '{sentence[:80]}'")
                         await self.realtime_client.chat_tts_text(
                             content=sentence,
                             start=not tts_started,
                             end=False,
                         )
+                        if not tts_started:
+                            await self._emit_status("speaking")
                         tts_started = True
+                        self._tts_segment_done = False
 
             # Handle remaining text in buffer (last sentence without ending punctuation)
             remaining = sentence_buf.strip()
             if remaining:
                 remaining = self._clean_markdown(remaining)
             if remaining:
+                # If previous TTS segment completed, close and wait before restart
+                if tts_started and self._tts_segment_done:
+                    logger.info(f"[Voice] TTS segment done, closing session and waiting for delivery (remaining)")
+                    await self.realtime_client.chat_tts_text(content="", start=False, end=True)
+                    self._tts_complete_event.clear()
+                    try:
+                        await asyncio.wait_for(self._tts_complete_event.wait(), timeout=10)
+                    except asyncio.TimeoutError:
+                        logger.warning(f"[Voice] Timeout waiting for TTS complete, proceeding anyway")
+                    tts_started = False
+                    self._tts_segment_done = False
+                    logger.info(f"[Voice] TTS delivery done, starting new session for remaining")
+
                 logger.info(f"[Voice] Sending TTS remaining: '{remaining[:80]}'")
                 await self.realtime_client.chat_tts_text(
                     content=remaining,
                     start=not tts_started,
                     end=False,
                 )
+                if not tts_started:
+                    await self._emit_status("speaking")
                 tts_started = True
+                self._tts_segment_done = False
 
             # Send TTS end signal
             if tts_started: