Speaker Diarization for overlapping audios

[Pranav Atre]

Hi, 

I am using pre-trained model http://kaldi-asr.org/models/m6 for speaker diarization and following this link for reference. While I am getting good results for audios where speakers don't overlap, I am not getting necessary results for audios where speakers are overlapping while speaking. Firstly, is it possible to get speaker labels correctly for such audio files? How would that output look like? or Should I continue checking for any issues in the script logic?

Attaching audio file and rttm file.

Thanks

[Desh Raj]

Kaldi doesn't support overlapping speaker diarization, meaning that it will only predict 1 speaker in the overlapping segments (and the predicted speaker is likely to be wrong since the x-vector extractor is not trained to deal with overlapping speech).

There is a ton of research trying to do overlap-aware diarization (see EEND or TS-VAD for example). If you want to do clustering based diarization with overlap assignment, you can use an external overlap detector and simple heuristics for overlap assignment. Some example recipes for doing this can be found on my repo here: 

https://github.com/desh2608/diarizer

Hope this helps.

Desh 

--
Go to http://kaldi-asr.org/forums.html to find out how to join the kaldi-help group
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/9025ed31-0414-47b2-8a8e-2202e8beb31an%40googlegroups.com.

[Pranav Atre]

Hi Desh,

Thank you so much for your reply. Will check this out.