[Feature] speaker identification or diarization

[mohit-levers]

Feature Description

What is the suggested way to identify speaker identification or diarization from the audios?

Use Case

We need to be able to identify speakers in multi turn audio recordings and also the timestamps.

[Emre-Akgul]

Use external models like nvidia-nemo or pyannote

[mohit-levers]

@Emre-Akgul Thanks for the suggestion.

Since pyannote has their own models, we want to evaluate omnilingual-asr models for our use case.

Could you please suggest what is best way to carry out this evaluation?

[urroxyz]

My PR (#22) will integrate timestamp extraction by default when committed.

I am working on my own project outside OmniASR that will allow for speaker verification through the native models with minimal training overhead, but that will take me some time to get running.

In any case, it is important to note that the models cannot diarize on their own. They just aren't trained for that. They are speaker-invariant.

@Emre-Akgul is right. You'll have to use additional models. I doubt the original researchers have any recommendations or specific evaluation requirements. Just create your own custom pipeline for your intended use case.