Speaker-blind speech recognition

[juanmc2005]

Depends on #143

Adding a streaming ASR pipeline needed a big refactoring (that began with #143).
This PR continues this effort to allow a new type of pipeline that transcribes speech instead of segmenting it.
A default ASR model based on Whisper is provided, but the dependency is not mandatory.

Additional modifications were also needed to make Whisper compatible with batched inference.
Note that we do not condition Whisper on previous transcriptions here. I expected this to degrade transcription quality but I found it rather robust in my experiments with the microphone and spontaneous speech in various languages (English, Spanish and French).

The new Transcription pipeline can also use a segmentation model as a local VAD to skip non-voiced chunks. In my experiments, this worked better and faster than using Whisper's no_speech_prob.

Transcription is also compatible with diart.stream, diart.benchmark, diart.tune and diart.serve (hence diart.client too).

Still missing

• README examples and possible restructuring

Changelog

TBD

[BlokusPokus]

Is this feature considered implemented?

[juanmc2005]

@BlokusPokus it seemed to work last time I tried but I didn't merge because I wanted to include a faster implementation of Whisper and I needed to clean up the code. Feel free to try it out but it's a pretty old version of the library. I need to find some time to update this PR. If you feel like it, it would be an amazing contribution!

[GeorgeDeac]

Yeah we definitely need a faster-whisper / WhisperLive implementation. WhisperLive also integrated VAD and I see it has some overlapping features.

[ywangwxd]

Yeah we definitely need a faster-whisper / WhisperLive implementation. WhisperLive also integrated VAD and I see it has some overlapping features.

Have you started to work on the faster-whisper implementation?
Is whisperX helpful to you? It uses faster-whisper.

[juanmc2005]

@GeorgeDeac @ywangwxd This feature is pretty much finished. But I think it would be nice to replace the default ASR by a faster one, like the ones you mention. Feel free to work on top of this PR, as I'm not too available to work on this as I'd like.
I'd gladly merge a PR to replace the default ASR and add a couple of examples in the README. I can take care of resolving the conflicts with the main branch after the feature is good

[arbuckle]

I'd suggest creating some basic tutorial content before merging into main - I tried this branch out and it was a struggle to integrate both Transcription and VAD into a single pipeline - something that's necessary when using Whisper in order to avoid hallucinated words in transcripts. (I get "Thank You" a lot when brief silences are transcribed by whisper)

I also ran into sampling issues when calling a transcription pipeline via diart.client. - these weren't present when doing simple diarization, and I'm not sure whether they're related to my own inexperience with this toolchain or suggestive of an issue with the code. In any case, a robust exemplar for a chained pipeline would be a good addition to the README. I created a couple:

simple: https://gist.github.com/arbuckle/d41cb5e25ccc588f4a98b8430eca40b5
complex: https://gist.github.com/arbuckle/5163d435ba174ee3ae866e789fa03f23

The complex example works, but it feels wrong and the performance is not yet to my satisfaction. (VAD needs to be tuned)