Can I change how many seconds the model looks at one time?

[jayavanth]

Basically, I want real-time STT with the medium model

I tried changing the CHUNK_SIZE in audio.py but it throws

Traceback (most recent call last):
  File "PYPATH/whisper/bin/whisper", line 8, in <module>
    sys.exit(cli())
  File "PYPATH/whisper/lib/python3.8/site-packages/whisper/transcribe.py", line 279, in cli
    result = transcribe(
  File "PYPATH/whisper/lib/python3.8/site-packages/whisper/transcribe.py", line 82, in transcribe
    _, probs = model.detect_language(segment)
  File "PYPATH/whisper/lib/python3.8/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context
    return func(*args, **kwargs)
  File "PYPATH/whisper/lib/python3.8/site-packages/whisper/decoding.py", line 43, in detect_language
    mel = model.encoder(mel)
  File "PYPATH/whisper/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl
    return forward_call(*input, **kwargs)
  File "PYPATH/whisper/lib/python3.8/site-packages/whisper/model.py", line 152, in forward
    assert x.shape[1:] == self.positional_embedding.shape, "incorrect audio shape"
AssertionError: incorrect audio shape

[jongwook]

The model can consume only 30 seconds at once, so the CHUNK_SIZE can't increase. model.transcribe() is designed to be used for audio files longer than 30 seconds, which works by successively transcribing audio using a 30s-long sliding window.

About the possibility of real-time transcription, please see my answers in #2.