revert back to using PyAV instead of torch audio

[MahmoudAshraf97]

This PR reverts the torchaudio code that was added in #856 and removes torchaudio dependency but still keeps torch

the reason why torch wasn't removed in this PR is that feature extraction still depends on it and I didn't want to include the numpy feature extraction in this PR to keep it simple and to reduce the number of conflicts to be resolved with #936

this should be merged before #936, after both are merged, a new PR will be created to completely remove the torch dependency

[joiemoie]

Will removing torch remove the supposed FFT speedup?

[MahmoudAshraf97]

Will removing torch remove the supposed FFT speedup?

Yes, but I have a new implementation using numpy in progress that is faster than the old implementation
These are the performance figures on my machine for a 30s segment:
Current Torch on cpu vs new numpy vs old numpy

I'm still investigating the possibility of GPU acceleration

[joiemoie]

Great to hear! I work with segments about 10 seconds long, so no benefit from batching. However, I am curious and possibly interested into bumping up to the latest commit because of this FFT speedup and especially interested in GPU acceleration

[MahmoudAshraf97]

these are the latest performance figures for the new FE @joiemoie

# torch on CPU
7.17 ms ± 119 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
# torch on GPU
797 µs ± 10.4 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)
# New numpy on CPU
14.6 ms ± 516 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
# New CuPY on GPU
1.73 ms ± 43.6 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)
# Old numpy on CPU
63.9 ms ± 2.73 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

[ozancaglayan]

Do you take into account the overhead of moving long audio tensors to GPU when doing the above measurements on GPU?

Update: I tried it and its around 1ms at most. My numpy variant is around ~30ms per 30sec as well.

[ozancaglayan]

Another option could actually be to move the feature extraction layer to CTranslate2 as well, whisper.cpp has some implementations
https://github.com/ggerganov/whisper.cpp/blob/master/src/whisper-mel-cuda.cu

[MahmoudAshraf97]

Another option could actually be to move the feature extraction layer to CTranslate2 as well, whisper.cpp has some implementations https://github.com/ggerganov/whisper.cpp/blob/master/src/whisper-mel-cuda.cu

OpenNMT/CTranslate2#1419 (comment)

[carolinaxxxxx]

@MahmoudAshraf97 good job sir. However I have a problem - it does not respect the --hotwords option at all in standard interference and batching mode. I tested on different materials. In the "old" version before the batching mode was introduced, --hotwords worked very well.

[Asmar097]

Good job