The transcribe functions still runs pad_or_trim on the mel spectrograms

[jpc]

There was a fix committed for padding the input audio to 30s multiples:
919a713

If I understand correctly, after the fix, there is no need to ever pad the mel spectrogram and yet the pad_or_trim is still called on mel spectrograms:

whisper/whisper/transcribe.py

Line 231 in e8622f9

mel_segment = pad_or_trim(mel_segment, N_FRAMES).to(model.device).to(dtype)

I don't think this causes any bugs right now (the pad_or_trim is always a no-op) but was quite misleading when reading the code and trying to understand how the padding works.