Use sliding window for realtime ASR preview

README.md

@@ -98,6 +98,7 @@ VOICEPRINT_STORE=/data/asr_voiceprints.json
 VOICEPRINT_MATCH_THRESHOLD=0.45
 REALTIME_PARTIAL_MIN_SECONDS=1.5
 REALTIME_PARTIAL_INTERVAL_SECONDS=2.0
+REALTIME_PARTIAL_WINDOW_SECONDS=6.0
 ```
 
 说明：

asr/main.py

@@ -103,6 +103,7 @@ VOICEPRINT_STORE = os.environ.get("VOICEPRINT_STORE", "/tmp/asr_voiceprints.json
 VOICEPRINT_MATCH_THRESHOLD = float(os.environ.get("VOICEPRINT_MATCH_THRESHOLD", "0.45"))
 REALTIME_PARTIAL_INTERVAL_SECONDS = float(os.environ.get("REALTIME_PARTIAL_INTERVAL_SECONDS", "2.0"))
 REALTIME_PARTIAL_MIN_SECONDS = float(os.environ.get("REALTIME_PARTIAL_MIN_SECONDS", "1.5"))
+REALTIME_PARTIAL_WINDOW_SECONDS = float(os.environ.get("REALTIME_PARTIAL_WINDOW_SECONDS", "6.0"))
 
 
 def _voiceprint_key(database_name: str, collection_name: str) -> str:
@@ -261,6 +262,12 @@ def _write_pcm_wav(audio_bytes: bytes) -> str:
     return tmp_path
 
 
+def _recent_pcm(audio_chunks: list[bytes], window_seconds: float) -> bytes:
+    max_bytes = max(int(window_seconds * 32000), 32000)
+    audio_bytes = b"".join(audio_chunks)
+    return audio_bytes[-max_bytes:]
+
+
 def _text_delta(previous: str, current: str) -> str:
     previous = previous or ""
     current = current or ""
@@ -298,7 +305,7 @@ async def run_asr_websocket(websocket: WebSocket):
                     last_partial_at = now
                     partial_path = None
                     try:
-                        partial_path = _write_pcm_wav(b"".join(audio_chunks))
+                        partial_path = _write_pcm_wav(_recent_pcm(audio_chunks, REALTIME_PARTIAL_WINDOW_SECONDS))
                         result = await run_asr(partial_path, language="zh", response_format="json")
                         partial_text = result.get("text", "") if isinstance(result, dict) else ""
                         if partial_text and partial_text != last_partial_text:
@@ -341,11 +348,14 @@ async def run_asr_websocket(websocket: WebSocket):
             if isinstance(result, dict):
                 text = result.get("text", "")
 
-        await websocket.send_json({
-            "mode": "2pass-offline",
-            "text": text,
-            "is_final": True,
-        })
+        try:
+            await websocket.send_json({
+                "mode": "2pass-offline",
+                "text": text,
+                "is_final": True,
+            })
+        except Exception as e:
+            logger.warning(f"Skip final transcript because websocket is closed: {e}")
     finally:
         if tmp_path and os.path.exists(tmp_path):
             os.unlink(tmp_path)

config.env

@@ -31,6 +31,7 @@ VOICEPRINT_STORE=/data/asr_voiceprints.json
 VOICEPRINT_MATCH_THRESHOLD=0.45
 REALTIME_PARTIAL_MIN_SECONDS=1.5
 REALTIME_PARTIAL_INTERVAL_SECONDS=2.0
+REALTIME_PARTIAL_WINDOW_SECONDS=6.0
 
 # fastapi_wss legacy/local ASR-with-speaker helper
 ASR_LOCAL_URL=http://asr-test:59805/asr