Error in language identification using whisper large

[Aafiya-H]

Hello,
While performing language identification on Fleurs, I receive the following error message:

RuntimeError                              Traceback (most recent call last)
[<ipython-input-10-3c79e22d458b>](https://localhost:8080/#) in <cell line: 6>()
     14     audio = whisper.pad_or_trim(audio)
     15     mel = whisper.log_mel_spectrogram(audio).to(model_large.device)
---> 16     _, probs = model_large.detect_language(mel)
     17     detected_language.append(max(probs, key=probs.get))
     18     # print(f"Detected language: {max(probs, key=probs.get)}")

9 frames
[/usr/local/lib/python3.10/dist-packages/torch/nn/modules/conv.py](https://localhost:8080/#) in _conv_forward(self, input, weight, bias)
    304                             weight, bias, self.stride,
    305                             _single(0), self.dilation, self.groups)
--> 306         return F.conv1d(input, weight, bias, self.stride,
    307                         self.padding, self.dilation, self.groups)
    308 

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

The code snippet works for other sizes of Whisper. Could you please guide me on how to fix this?

[glangford]

From the announcement of large-v3,

"The input uses 128 Mel frequency bins instead of 80" ...
#1762

By default calling log_mel_spectrogram directly without extra parameters will use 80 and from the error message I suspect this is the problem. You can override this when calling the low level API, here is the function signature:

whisper/whisper/audio.py

Line 110 in ba3f3cd

def log_mel_spectrogram(