Conversation
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
|
Gentle bump. Long quiet STT sessions still drop on idle websocket timeouts. This enables aiohttp PING/PONG heartbeat only for the inference STT streams, so normal quiet-user sessions are completely untouched (no application-level pings). Happy to make it configurable if you'd rather gate it. |
| f"{base_url}/stt?model={self._opts.model}", headers=headers | ||
| f"{base_url}/stt?model={self._opts.model}", | ||
| headers=headers, | ||
| heartbeat=WS_HEARTBEAT, |
There was a problem hiding this comment.
this only check the connection between client to the gateway, but cannot fix the issue described in the original issue
If the gateway WebSocket connects successfully but the upstream STT provider (Deepgram, etc.) silently stops returning transcripts — e.g., the gateway-to-provider tunnel goes flaky, or TCP goes half-open — the recv loop will block indefinitely.
|
longcw's review pointed out the heartbeat covers client↔gateway, not the gateway→provider silence from the original issue; I have not reworked it in 59 days. Closing until a real fix exists. |
2 participants
Summary
Fixes #5717.
To verify
python -m py_compile livekit-agents\livekit\agents\inference\stt.py tests\test_inference_stt_fallback.pyuv run ruff check livekit-agents\livekit\agents\inference\stt.py tests\test_inference_stt_fallback.pyuv run ruff format --check livekit-agents\livekit\agents\inference\stt.py tests\test_inference_stt_fallback.pyuv run pytest tests\test_inference_stt_fallback.py -q