Large Model Issue - Same Line Inserted for All Detected Speech

[ehobson]

I've encountered an issue with the Large model in both the Python environment and the main environment, using both translation and transcription. It will translate or transcribe only one sample of detected speech and then insert that one translation/transcription for all speech detected in the file. Example:

      [01:30.000 --> 01:54.000]  So, how are you today?
      [01:54.000 --> 02:17.000]  How's the weather?
      [02:17.000 --> 02:44.000]  How's the weather?
      [02:44.000 --> 03:12.000]  How's the weather?
      [03:12.000 --> 03:39.000]  How's the weather?
      [03:39.000 --> 04:06.000]  How's the weather?
      [04:06.000 --> 04:34.000]  How's the weather?
      [04:34.000 --> 05:02.000]  How's the weather?
      [05:02.000 --> 05:25.000]  How's the weather?
      [05:25.000 --> 05:53.000]  How's the weather?
      [05:53.000 --> 06:22.000]  How's the weather?

The one or two translations/transcriptions that it identifies are usually very accurate, but it fails to process any other speech in the file. This error has happened for all filetypes, sample rates, and bitrates used.

Edit: It's not always the second line of detected speech that it gets stuck on. Sometimes all translations/transcriptions will be the same except for the fifth, seventh, tenth, or some other random line of speech.

[glangford]

There is a suggestion in #192 to try and address this known failure mode. See also the discussion in #112.

[crecos]

The solution in #192 didn't help me personally.

The hallucination issue is currently being addressed in #679