Speker Diarization Using Whisper Transcription and Nvidia NeMo

[MahmoudAshraf97]

Hello, I've built a pipeline Here to enable speaker diarization using whisper's transcriptions. It includes preprocessing that separates the vocals from other sounds, and post processing by realigning the transcriptions according to punctuations (thanks to @mu4farooqi). It also uses WhisperX (by @m-bain) for timestamp correction.

From my trials, the results are better than the PyAnnote approach mentioned in #264
The code is originally written to handle 1hr+ podcasts so no need to split the audio in advance
Feel free to try it and give me your feedback and suggestions

[Jeronymous]

Cool, that's a great contribution!

FYI, there are other approaches than WhisperX to get word timestamps, that do not require an additional wav2vec model.
I mean approach based on Whisper's cross-attention weights.
In particular, it's much easier to support several languages with such approaches.
I did an implementation of that here.

[magicniuniu]

I have succeeded, whisper timestamped 2.wav --model tiny --output_dir d:, this is what is needed in reality, thank you Jeronymous for the py

[MahmoudAshraf97]

@Jeronymous I tried whisper-timestamped but I faced some drawbacks that prevented the replacement of WhisperX

1. Strangely, it had higher VRAM usage that whisper+wav2vec combined, that was observed on different machines
2. it uses the timestamps generated by whisper, from my testing, these timestamps were found to be inaccurate by more than 30 seconds in long audio files such as podcasts and slightly inaccurate in small files, this can be tolerated in case of captioning but in diarization, we rely on the sync between the timing generated by wav2vec and the timing generated by NeMo, thus the timestamps affect diarization quality drastically

[m-bain]

@MahmoudAshraf97 I am finding NeMo is very slow to run compared to pyannote -- even with cuda, do you have any idea on how to speed this up?

[Jeronymous]

Thanks a lot for your feedback @MahmoudAshraf97
I'll investigate how to improve VRAM/RAM usage.

these timestamps were found to be inaccurate by more than 30 seconds

More than 30 seconds is very weird as whisper is splitting the audio in chunks of 30 seconds at most.
So it sounds like a bug somewhere... Did you see these unexpected timestamps when running whisper (not whisper-timestamped)?

[MahmoudAshraf97]

@m-bain I think it's due to the underlying NeMo structure, if someone can extract the underlying pytorch models and handle the inference it can be a bit faster, also note that the NeMo part is totally independent from whisper so it's completely parallelizable if enough VRAM exists

@Jeronymous Yes, the incorrect timestamps are also a problem with whisper, that's why it's not entirely an alignment problem, we need to regenerate accurate timestamps from an external source

[Jeronymous]

I'll investigate how to improve VRAM/RAM usage.

On the last version of whisper-timestamped, there is no problem of VRAM/RAM.
The memory usage is comparable to WhisperX, and there is little overhead in comparison to standard Whisper.

I also benchmarked the accuracy of word timestamp on a French dataset for which I had an accurate annotation of timestamps, and the performance of WhisperX and whisper-timestamped are very close.
Just whisper-timestamped is able to locate a little more things, like digits ("32") and symbols ("€", "%"...).

I will soon implement an approach that uses VAD to be independent of whisper timestamp prediction.
And also some heuristics to improve things around disfluencies that are not transcribed by Whisper (they are currently a problem both for WhisperX and whisper-timestamped).

[josephcappadona]

Hey, this worked pretty well! The diarization isn't the best, but it's pretty good and should be good enough for many use cases. Thanks for sharing!

[Majdoddin]

www.lexicaps.com seamlessly adds diarization to Whispers transcription. No 3rd party packages.
Announcement: #1537
Repo: https://github.com/Majdoddin/lexicaps

[chainyo]

Hi, we are also building an ASR tool using Whisper and NVIDIA NeMo for diarization on public: https://github.com/Wordcab/wordcab-transcribe

Audio file transcription will never be a struggle anymore, plus we provide a top-class API on top of the full process.

[dgoryeo]

Thanks @chainyo . Out of curiosity, was there any particular reason you selected Nemo for diarization? I see quite a few pyannote implementations but less Nemo. I was wondering.

[chainyo]

Good question. In my experience, NeMo diarization (not the entire MSDD pipeline), based on a multiscale auto-tuning spectral clustering, is much more resilient than the pyannote version.

[Arno-Riverside]

Hi @chainyo, what do you mean by "more resilient" ? I'm currently comparing pyannote vs nemo on my dataset and I'm struggling to tune Nemo's hyperparameters to outperform pyannote.

[chainyo]

Hi @chainyo, what do you mean by "more resilient" ? I'm currently comparing pyannote vs nemo on my dataset and I'm struggling to tune Nemo's hyperparameters to outperform pyannote.

From what I tested some months ago, the accuracy/precision of the returned timestamps were way better than pyannote.

[tristancgardner]

Same here in 2024

[piegu]

Hi.
The code of @MahmoudAshraf97 (Whisper+Nemo for diarization) is great. Thank you for that.
Now, I'm looking for an API that uses it for easy integration in an app.
Who knows online services about Whisper+Diarization?
Can you share links here?
Thank you.
Pierre