Currently training LibriSpeech/ASR/transducer/ with triangularclassic and without to see if it works. I don't expect the resulting WER difference to be any significant past the margin of error, but it is worth checking.

I am not sure if I would want to keep the bin alignment option; kaldi/k2 doesn't seem to be doing it and it doesn't feel like it would be beneficial especially without high n_fft.

It took us a long time due to many high priority things, and mostly because it does not seem to affect the results, but this is an important fix. Maybe @idruker-cerence wants to have a look.

FWIW I have a stock RNN-T LibriSpeech Conformer being trained again with the triangularclassic option and at epoch ~34 there doesn't seem to be much of a difference yet.
Will report on the results when actually finished.

It took us a long time due to many high priority things, and mostly because it does not seem to affect the results, but this is an important fix. Maybe @idruker-cerence wants to have a look.

Hello

I am not a python expert so it's hard to me to review the code. However, I have a C++ implementation. We can do the following. You calculate the mel-spectrum with your new function and give me the numbers. Then I verify them with my C++ function.

FWIW I have a stock RNN-T LibriSpeech Conformer being trained again with the triangularclassic option and at epoch ~34 there doesn't seem to be much of a difference yet. Will report on the results when actually finished.

Current code, lines are respectively test-clean and test-other:

Epoch loaded: 100 - test loss: 1.80e-02, test CER: 8.81e-01, test WER: 2.57
Epoch loaded: 100 - test loss: 2.16e-02, test CER: 2.50, test WER: 5.88

New "triangularclassic":

Epoch loaded: 100 - test loss: 1.88e-02, test CER: 8.57e-01, test WER: 2.59
Epoch loaded: 100 - test loss: 2.12e-02, test CER: 2.56, test WER: 5.97

That's only one run with 100 epochs at that, but it doesn't seem like a significant accuracy difference (~1% relative difference for only one run).

As for correctness, from my testing, "triangularclassic" matches what torchaudio produces, which itself is adapted from kaldi (which I also checked against, or rather k2, which reuses a good chunk of the same logic). So I think it would be ok to add for correctness.

I am tempted to turn this PR into only adding the "correct" filter as an option, but keep the default the same, and remove the snap_to_bins (which I don't think matters, and that I didn't test extensively).

At least in speech processing, correct me if I'm wrong, it doesn't seem so common to round the center frequencies to the closest bin in definitions of the mel filterbank (which is what I tried to implement as snap_to_bin).
I think you can only get "clean-looking" filter banks as usually shown in blog articles/papers if you really crank up the FFT resolution, because otherwise you're either trading off between obviously distorted central frequency spacing (due to the rounding) or not having the central frequency have a peak of 1.0 for on a specific FFT bin.