wq2012/awesome-diarization

A curated list of awesome Speaker Diarization papers, libraries, datasets, and other resources.

awesome-diarization's Issues

Question about training chinese speech diarization model

It was memtioned in your paper that google used 36M utterance and 18K speaker to train the model. If I want to training a chinese version of the speech diazarition model. How much data do I need for that?
Thank you.

Leadboard

Where has the leaderboard gone?

Multi-scale Speaker Diarization with Dynamic Scale Weighting

Do you think the method from Nvidia in 2022 is impact:

Multi-scale Speaker Diarization with Dynamic Scale Weighting : https://arxiv.org/pdf/2203.15974.pdf

And

TitaNet: Neural Model for speaker representation with 1D Depth-wise separable convolutions and global context

I found it in the leaderboard of Paper with code

nanodrz

Not a publication but I'll be working on getting this more e2e approach working with transformers:

https://github.com/mogwai/nanodrz

How to use AMI dataset to evaluate the DER performance?

Hi, I have seen some authors use AMI corpus to make evaluation on diarization task. But there is no more details about how to evaluate specifically. Like how to choose the dev and test part of the AMI,and how to make the corresponding data preparation.
Is there any guidance about using AMI corpus to evaluate the task？
Thanks.
@wq2012

Leaderboard / benchmark ?

It would be nice to also have some kind of leaderboard associated to each of the listed datasets.

One (bad) way of doing it would be to use numbers reported in papers. It is bad for several reasons:

• one cannot be sure the exact same evaluation protocol was used
• one cannot be sure the exact same metric was used

A better way of doing it would be to ask authors to provide their output files and run the evaluation for them (possibly automatically on each pull request) but this does not solve the first problem. We could use pyannote.metrics for that.

An even better way would be to ask them to provide runnable pre-trained systems and run them for them but this would need a lot of work to ask from the authors and to setup.

An utopian way would to ask them to provide trainable systems.

Anyway, maybe it is too much to ask and existing challenges like DIHARD and Albayzin are probably enough...