Add new augmentations: lowpass filtering, lossy compression (opus/mp3/vorbis)

[racoiaws]

This PR adds two operations on Cuts/Recordings and corresponding randomized CutSet transforms

1. Lowpass filtering
2. In-memory lossy compression (with immediate decoding back to raw waveform)

Both would be useful for training robust ASR models, speech enhancement models, etc.

TODO

• tests
• docstrings for new methods in lhotse.audio.recording.Recording

[pzelasko]

When you write OPUS files with soundfile, it adds extra information in the file header about the original sampling rate, so that when you load this file with soundfile later, it's resampled from 48k before returning the audio array (unlike most other tools).

[racoiaws]

Nice, did not know that

[racoiaws]

Noted

[pzelasko]

Since scipy is an optional lhotse dependency, please move this import inside __call__, and add an import guard (from lhotse.utils) before it:

if not is_module_available("scipy"):
    raise ImportError("In order to use Lowpass transforms, run 'pip install scipy'")

[racoiaws]

Done

[pzelasko]

Can you also move all soundfile imports to local function scopes? IIRC importing this globally used to silently break documentation builds.

[racoiaws]

Done

[pzelasko]

Great work! Please also add unit tests for calling transform methods on Recording and Cut, and for calling cut_transforms on CutSet

[racoiaws]

Thanks for the review, addressed comments

Will add tests a bit later