HLS Transcription Stops After A Couple Minutes

[austinm1120]

I set up a local HLS stream playing a long video of someone talking.

Everything seems great until after exactly 2 minutes in the transcription stops completely.

INFO:faster_whisper:Processing audio with duration 00:07.936
INFO:faster_whisper:Processing audio with duration 00:02.984
INFO:faster_whisper:Processing audio with duration 00:03.032
INFO:faster_whisper:Processing audio with duration 00:01.432
INFO:faster_whisper:Processing audio with duration 00:03.480
INFO:faster_whisper:Processing audio with duration 00:05.272
INFO:faster_whisper:Processing audio with duration 00:01.152
INFO:faster_whisper:Processing audio with duration 00:03.200
INFO:faster_whisper:Processing audio with duration 00:02.548
INFO:faster_whisper:Processing audio with duration 00:04.596
INFO:faster_whisper:Processing audio with duration 00:01.796
INFO:faster_whisper:Processing audio with duration 00:03.844
INFO:faster_whisper:Processing audio with duration 00:02.484
INFO:faster_whisper:Processing audio with duration 00:02.484
INFO:faster_whisper:Processing audio with duration 00:02.484
INFO:faster_whisper:Processing audio with duration 00:02.484
INFO:faster_whisper:Processing audio with duration 00:02.484
INFO:faster_whisper:Processing audio with duration 00:02.484
INFO:faster_whisper:Processing audio with duration 00:02.484

In the server logs i can see that chunks of variable length are processed by the server. However the problem starts when the "00:02.484" chunks keep getting processed. I'm unsure if its just continuing to send the same chunk and it keeps translating it therefore the client appears to be "stuck" or if its stuck in a different loop of some sort.

Setting use_vad to True doesn't seem to make a difference.

I have tried both on Mac (M3 Max chip) and Windows 10. Both docker and python server. Both produce the same results.

This is the client code:

from whisper_live.client import TranscriptionClient
client = TranscriptionClient(
  "localhost",
  9090,
  lang="en",
  translate=False,
  model="small",
  use_vad=False,
)


client(hls_url="http://localhost.m3u8")

[makaveli10]

@austinm1120 Thanks for reporting the issue, does whisper-live behave similarly with other input types as well or is it only HLS?

[nums]

Same issue with rtmp / rtsp

[aliuspetraska]

same happens to me, but this time it's exactly 10 minutes every time:

[INFO]: Server disconnected due to overtime.
[INFO]: Websocket connection closed: 1000: 

Investigating on my own and if I will come to any fixes, will let you know.

[aliuspetraska]

OK, I found the "issue" :) It's not an issue, it's was designed to work that way: https://github.com/collabora/WhisperLive/blob/main/whisper_live/server.py#L28

So in my case I'll do refactoring to strip that part.