Awsaf
> This appears to be correct and normal: [pytorch/audio#1058](https://github.com/pytorch/audio/issues/1058) > > > I found librosa use slaney normalization for the mel-filterbank creation as the default, while torchaudio is no normalization...
@fchollet We can use an external mel matrix like this if we want to use `norm="slaney"`. I've checked the result, and it matches with `librosa`, which means a mismatch is...
@fchollet should I add `correctness` test with `torchaudio`? Also do we need `tf_data` compatibility? I tried adding `tf_data_compatibility` test. But it fails. **code**: ```py input_data = np.random.random((2, 8000)) layer =...
> As for whether we should do it -- how do you think the layer will be used? Inside a model? Externally? Using this layer in a model is definitely...
@fchollet I've added support for `tf.data` but as it required using `self.backend` instead of `keras.ops`, I had to move the independent functions into `MelSpectrogram` layer. I've checked it, it runs...
@fchollet I've aligned the `crop_images` to `pad_images` with similar arguments. Kindly let me know your feedback.
Thanks @george-gca for the suggestion, I will try it out and let you know.
cc: @fchollet
Yes. `layers/preprocessing` would be nice. I will create the PR soon with previous progress.
@fchollet , I tried to implement the `linear_to_mel_weight_matrix` in `keras_core` for `MelSpectrogram` layer, and during this, I came across a few other dependencies that seem to be missing in `keras_core`....