Removes use of torchaudio and moves transforms inside of NeMo

[blisc]

Important

The Update branch button must only be pressed in very rare occassions.
An outdated branch is never blocking the merge of a PR.
Please reach out to the automation team before pressing that button.

What does this PR do ?

Removes use of torchaudio.transforms and moves transforms inside of NeMo.
NOTE: we will use torchsquirm in nemo/collections/audio/metrics/squim.py and nemo/collections/tts/models/magpietts_preference_optimization.py

Collection: audio, asr, tts

Changelog

• Move frequently used torchaudio transform into NeMo

PR Type:

• New Feature
• Bugfix
• Documentation

[nithinraok]

LGTM.

@chtruong814 / @ko3n1g review for docker related changes.

[MahmoudAshraf97]

This PR is a breaking change to older models, please take action before it makes it to the next release

RuntimeError: Error(s) in loading state_dict for EncDecCTCModelBPE:
Missing key(s) in state_dict: "preprocessor.featurizer.window", "preprocessor.featurizer.fb".
Unexpected key(s) in state_dict: "preprocessor.featurizer._mel_spec_extractor.spectrogram.window", "preprocessor.featurizer._mel_spec_extractor.mel_scale.fb".

[pzelasko]

Which models is it breaking / how old?

[MahmoudAshraf97]

I managed to reproduce it with a model trained using v1.23.0, I have another model that was trained using v2.2.1 that did not reproduce the issue, these are internal models that I cannot share but I'm happy to test models published on HF or prepare a minimum repro if needed

[MahmoudAshraf97]

Further investigation shows that this is reproducible with any model that was trained with preprocessor.use_torchaudio=True regardless of the version used to train it

[pzelasko]

Torchaudio was removed as a dependency. Can you migrate all models to non torchaudio preprocessor?

[MahmoudAshraf97]

I don't mind doing that, in fact I stopped using it a while ago, the problem arises when we try to load models that were trained using torchaudio in the preprocessor and that fails, the solution imo would be having a translation code to match the key names in the state dict during the model loading process or a script to convert old .nemo files that used torchaudio to a format that is accepted by the new versions (just modify the parameter names in the state dict)

[pzelasko]

@MahmoudAshraf97 see if this helps #15437