Improve --model argument handling and help message

[spartanhaden]

This PR introduces the following updates to the whisper/transcribe.py script:

• Enhancement of the --model argument handling and help message: The --model argument now provides a list of available model choices along with the default option when the --help flag is used. This enhances user experience by providing immediate visibility of the available options.

  • Previous message: --model MODEL name of the Whisper model to use (default: small)
  • Updated message: --model MODEL name of the Whisper model to use. Available models are: tiny.en, tiny, base.en, base, small.en, small, medium.en, medium, large-v1, large-v2, large-v3, large. You can also specify a path to a model checkpoint. (default: small)
  • Note: The choices=available_models() option was not used to allow the use of custom model checkpoints.

• Improved error message for incorrect model names: If a non-existing model name is used, the error message now functions as intended and indicates the error and provides the list of valid model names.

  • Previous message: whisper: error: argument --model: invalid valid_model_name value: 'some_incorrect_model_name'
  • Updated message: whisper: error: argument --model: model should be one of ['tiny.en', 'tiny', 'base.en', 'base', 'small.en', 'small', 'medium.en', 'medium', 'large-v1', 'large-v2', 'large-v3', 'large'] or path to a model checkpoint

[MohamedAliRashad]

large-v3 is not working, it's giving me this error.

RuntimeError: Given groups=1, weight of size [1280, 128, 3], expected input[1, 80, 3000] to have 128 channels, but got 80 channels instead

[FurkanGozukara]

large-v3 is not working, it's giving me this error.

RuntimeError: Given groups=1, weight of size [1280, 128, 3], expected input[1, 80, 3000] to have 128 channels, but got 80 channels instead

i tested yesterday worked very well

i am using python 3.10.11

[MohamedAliRashad]

@FurkanGozukara
It turned out that the function log_mel_spectrogram requires n_mels to be set to 128 because large-v3 works with 128 not 80 like large-v2.

[FurkanGozukara]

log_mel_spectrogram

what does it do?

[ihmily]

log_mel_spectrogram

what does it do?

I also encountered this issue, and the error message is Given groups=1, weight of size [1280, 128, 3], expected input[1, 80, 3000] to have 128 channels, but got 80 channels instead.

The solution is to modify the following line of code:

mel = whisper.log_mel_spectrogram(audio).to(model.device)

to:

mel = whisper.log_mel_spectrogram(audio, n_mels=128).to(model.device)

By explicitly setting n_mels=128, it might resolve the issue and allow the code to run properly. If it still doesn't work, you can try changing n_mels = 128 back to n_mels = 80.

In general, when higher frequency resolution is needed, selecting n_mels = 128 is recommended. A higher value of n_mels provides more Mel frequency filters, capturing more details and frequency components in the spectrogram representation.