Fix Gradio duplex audio playback clearing

api/go_server.py

@@ -195,7 +195,7 @@ def _merge_wavs_to_base64(wav_dir: str, wav_files: list) -> str:
     return ""
 
 
-def _wait_for_tts_files(tts_wav_dir: str, timeout_s: float = 15.0) -> list:
+def _wait_for_tts_files(tts_wav_dir: str, timeout_s: float = 45.0) -> list:
     """Wait briefly because llama.cpp-omni may return before TTS wav files finish."""
     deadline = time.time() + timeout_s
     last_files = []
@@ -217,7 +217,7 @@ def _wait_for_tts_files(tts_wav_dir: str, timeout_s: float = 15.0) -> list:
                     last_files = wav_files
                 if os.path.exists(os.path.join(tts_wav_dir, "generation_done.flag")):
                     return wav_files
-                if stable_seen >= 2:
+                if stable_seen >= 12:
                     return wav_files
         time.sleep(0.25)
     return last_files
@@ -385,6 +385,7 @@ async def voice_chat(req: VoiceChatRequest):
     wav_files = _wait_for_tts_files(tts_wav_dir)
     if wav_files:
         tts_audio_base64 = _merge_wavs_to_base64(tts_wav_dir, wav_files)
+        logger.info(f"TTS wav 合并完成: {len(wav_files)} 个片段")
     else:
         logger.warning(f"TTS wav 未在超时内生成: {tts_wav_dir}")
 
@@ -396,6 +397,7 @@ async def voice_chat(req: VoiceChatRequest):
         "text": text_output,
         "audio_base64": tts_audio_base64,
         "audio_sample_rate": 24000,  # TTS 默认采样率
+        "audio_files": len(wav_files),
         "timing": {
             "audio_prefill_ms": 0,
             "decode_ms": 0,

api/voice_helper.py

@@ -102,6 +102,7 @@ def chat_voice(audio_path: str) -> dict:
                 "text": data.get("text", ""),
                 "audio_base64": data.get("audio_base64", ""),
                 "round": data.get("round", 0),
+                "audio_files": data.get("audio_files", 0),
             }
         else:
             # fallback: 直接用文本对话（避免服务中断）

research/task_manager/pregopal_gradio_duplex_task.json

@@ -4,8 +4,8 @@
     {
       "stepId": "1",
       "description": "Main Task",
-      "status": "running",
-      "resultNote": "",
+      "status": "completed",
+      "resultNote": "Auto-completed: all sub-steps done.",
       "subSteps": [
         {
           "stepId": "1.1",
@@ -28,8 +28,8 @@
         {
           "stepId": "1.3",
           "description": "Verify Gradio microphone streaming VAD loop",
-          "status": "running",
-          "resultNote": "Backend verified; next validate Gradio microphone streaming and VAD loop.\nFunction-level VAD with synthetic tone triggered backend and appended chat text, but no audio path; next test uses real WAV chunks.\nAdded TTS merge cap to avoid abnormal huge wav_* outputs from non-speech/synthetic inputs during Gradio demo.",
+          "status": "completed",
+          "resultNote": "Function-level Gradio stream path passed: real WAV chunks triggered backend, produced temp wav, and later silence frames now return gr.update() instead of clearing audio.",
           "subSteps": [],
           "workDir": "C:\\Users\\Andre\\codes\\LJB\\hackthon\\for_codex\\PregoPal\\research\\task_manager\\workspace\\1\\3",
           "instructions": ""
@@ -37,8 +37,8 @@
         {
           "stepId": "1.4",
           "description": "Apply minimal demo UI or logging fixes only if needed",
-          "status": "pending",
-          "resultNote": "",
+          "status": "completed",
+          "resultNote": "Restarted Gradio PID 64656; HTTP 7889 returns 200. PregoAPI remains healthy on 8090.",
           "subSteps": [],
           "workDir": "C:\\Users\\Andre\\codes\\LJB\\hackthon\\for_codex\\PregoPal\\research\\task_manager\\workspace\\1\\4",
           "instructions": ""

ui/app_builder.py

@@ -81,6 +81,11 @@ def _reset_duplex_buffers():
     _SILENT_SECONDS = 0.0
 
 
+def _keep_audio():
+    """Keep the current Gradio audio value instead of clearing playback."""
+    return gr.update()
+
+
 def toggle_duplex(*args):
     """点按切换全双工对话状态"""
     global _DUPLEX_ACTIVE
@@ -94,7 +99,7 @@ def handle_stream_chunk(audio_chunk, chat_history):
     """处理 streaming 音频块: VAD -> 缓冲 -> 静音后调后端"""
     global _DUPLEX_ACTIVE, _AUDIO_BUF, _AUDIO_SR, _SPEECH_SECONDS, _SILENT_SECONDS, _LAST_CHUNK_LOG
     if not _DUPLEX_ACTIVE:
-        return chat_history, _thinking_html("✅ 已就绪"), None
+        return chat_history, _thinking_html("✅ 已就绪"), _keep_audio()
     if not chat_history:
         chat_history = []
 
@@ -103,7 +108,7 @@ def handle_stream_chunk(audio_chunk, chat_history):
     if audio_chunk is not None and isinstance(audio_chunk, tuple) and len(audio_chunk) == 2:
         sr, arr = audio_chunk
         if arr is None:
-            return chat_history, _thinking_html("🎤 聆听中..."), None
+            return chat_history, _thinking_html("🎤 聆听中..."), _keep_audio()
         if arr.dtype.kind == 'i':
             audio_np = arr.astype(np.float32) / 32768.0
         else:
@@ -111,7 +116,7 @@ def handle_stream_chunk(audio_chunk, chat_history):
         if len(audio_np.shape) > 1:
             audio_np = audio_np.mean(axis=1)
     if audio_np is None or len(audio_np) == 0:
-        return chat_history, _thinking_html("🎤 聆听中..."), None
+        return chat_history, _thinking_html("🎤 聆听中..."), _keep_audio()
 
     _AUDIO_SR = int(sr or _TARGET_SR)
     chunk_seconds = len(audio_np) / max(float(_AUDIO_SR), 1.0)
@@ -136,7 +141,7 @@ def handle_stream_chunk(audio_chunk, chat_history):
             )
             _reset_duplex_buffers()
             return _call_duplex_backend(full, chat_history, sr=buffered_sr)
-        return chat_history, _thinking_html("🎤 聆听中..."), None
+        return chat_history, _thinking_html("🎤 聆听中..."), _keep_audio()
 
     _SPEECH_SECONDS += chunk_seconds
     _SILENT_SECONDS = 0.0
@@ -148,7 +153,7 @@ def handle_stream_chunk(audio_chunk, chat_history):
         print(f"[PregoPal Duplex] max utterance reached dur={total_seconds:.2f}s", flush=True)
         _reset_duplex_buffers()
         return _call_duplex_backend(full, chat_history, sr=buffered_sr)
-    return chat_history, _thinking_html("🔊 正在听..."), None
+    return chat_history, _thinking_html("🔊 正在听..."), _keep_audio()
 
 
 def _call_duplex_backend(audio_np, chat_history, sr=_TARGET_SR):
@@ -172,6 +177,11 @@ def _call_duplex_backend(audio_np, chat_history, sr=_TARGET_SR):
             try: os.remove(path)
             except: pass
         if result.get("success"):
+            print(
+                f"[PregoPal Duplex] backend success round={result.get('round')} "
+                f"audio_files={result.get('audio_files')} audio_b64_len={len(result.get('audio_base64') or '')}",
+                flush=True,
+            )
             ai_text = result.get("text", "").strip()
             if not ai_text or len(ai_text) < 2:
                 sp = "你是PregoPal孕期营养健康顾问。请用中文简短回答。"
@@ -195,12 +205,12 @@ def _call_duplex_backend(audio_np, chat_history, sr=_TARGET_SR):
                 with open(out_path, "wb") as f:
                     f.write(base64.b64decode(result["audio_base64"]))
                 return (chat_history, _thinking_html("🎤 聆听中..."), out_path)
-            return chat_history, _thinking_html("🎤 聆听中..."), None
+            return chat_history, _thinking_html("🎤 聆听中..."), _keep_audio()
         else:
-            return chat_history, _thinking_html("🎤 聆听中..."), None
+            return chat_history, _thinking_html("🎤 聆听中..."), _keep_audio()
     except Exception as e:
         traceback.print_exc()
-        return chat_history, _thinking_html(f"错误: {str(e)[:50]}"), None
+        return chat_history, _thinking_html(f"错误: {str(e)[:50]}"), _keep_audio()
 
 
 def _thinking_html(text: str) -> str: