Add speaker-aware transcription

[juanmc2005]

Depends on #144

This PR adds a new SpeakerAwareTranscription pipeline that combines streaming diarization and streaming transcription to determine "who says what" in a live conversation. By default, this is shown as colored words in the terminal.

The feature works as expected with diart.stream and diart.serve/diart.client.
The main thing preventing full compatibility with diart.benchmark and diart.tune is the evaluation metric.
Since the output of the pipeline is annotated text with the format: [speaker0]Hello [speaker1]Hi, the metric diart.metrics.WordErrorRate will count labels as insertion errors.

Next steps: implement a SpeakerWordErrorRate that computes the (weighted?) average WER across speakers.

Changelog

TBD

[C0RE1312]

Hey, I am unable to use this:
(diart) :~/live-transcript$ diart.stream output.wav --pipeline SpeakerAwareTranscription
Traceback (most recent call last):
File "/home/user/anaconda3/envs/diart/lib/python3.8/site-packages/rx/core/operators/map.py", line 37, in on_next
result = _mapper(value)
File "/home/user/anaconda3/envs/diart/lib/python3.8/site-packages/diart/pipelines/speaker_transcription.py", line 325, in call
asr_outputs = self.asr(batch[has_voice])
File "/home/user/anaconda3/envs/diart/lib/python3.8/site-packages/diart/blocks/asr.py", line 65, in call
output = self.model(wave.to(self.device))
File "/home/user/anaconda3/envs/diart/lib/python3.8/site-packages/diart/models.py", line 80, in call
return super().call(*args, **kwargs)
File "/home/user/anaconda3/envs/diart/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1194, in _call_impl
return forward_call(*input, **kwargs)
File "/home/user/anaconda3/envs/diart/lib/python3.8/site-packages/diart/models.py", line 485, in forward
batch = whisper.log_mel_spectrogram(batch)
File "/home/user/anaconda3/envs/diart/lib/python3.8/site-packages/whisper/audio.py", line 148, in log_mel_spectrogram
stft = torch.stft(audio, N_FFT, HOP_LENGTH, window=window, return_complex=True)
File "/home/user/anaconda3/envs/diart/lib/python3.8/site-packages/torch/functional.py", line 632, in stft
return _VF.stft(input, n_fft, hop_length, win_length, window, # type: ignore[attr-defined]
RuntimeError: cuFFT error: CUFFT_INTERNAL_ERROR

[juanmc2005]

@C0RE1312 sounds like a problem with pytorch not being able to compute the FFT. Have you tried updating the dependenciesof both torch and whisper? it's a pretty old PR

[ywangwxd]

@C0RE1312 sounds like a problem with pytorch not being able to compute the FFT. Have you tried updating the dependenciesof both torch and whisper? it's a pretty old PR

Will this work with faster-whisper or any other faster version of whisper?

[ywangwxd]

@C0RE1312 sounds like a problem with pytorch not being able to compute the FFT. Have you tried updating the dependenciesof both torch and whisper? it's a pretty old PR

BTW, I noticed that the last commit was in the April of 2023. So this feature has no new commits for more than one year. Do this mean the feature implementation has finished but it was not merged into the main branch? I noticed in the readme page of this project, there was a note stating that this feature was comming soon but ready.

[juanmc2005]

@ywangwxd unfortunately I haven't had the time to work on this as I'd like. I prioritized other things like documentation and testing for #98
This should work okay if we update the dependencies and resolve conflicts with the main branch, but I think it's better to do a full rework because it's not very efficient. To begin with, it's using OpenAI's whisper, while there are faster implementations out there.
Please feel free to improve and work on top of this branch and open your own transcription PR, I can take a look regularly and guide you if you want to contribute.