Project: re-train ECAPA-TDNN on adults' and children's voices

sensein / senselab

senselab is a Python package that simplifies building pipelines for biometric (e.g. speech, voice, video, etc) analysis.

http://sensein.group/senselab/

Apache License 2.0

10 stars 3 forks source link

Project: re-train ECAPA-TDNN on adults' and children's voices #45

Open fabiocat93 opened 5 months ago

fabiocat93 commented 5 months ago

Description

Train ECAPA-TDNN on adults' and children's voices

Tasks

[ ] explore the impact of training data window on speaker diarization performance (in a child-adult setting) and on speaker verification
[ ] explore the impact of including children's data on speaker diarization performance (in a child-adult setting) and on speaker verification

Freeform Notes

No response

fabiocat93 commented 3 months ago

https://docs.google.com/document/d/1wGjFpzVY6hswNJXwLfep5LmpJi2sWnPUI4-_XOlTewI/edit

fabiocat93 commented 1 week ago

@900miles, do you happen to have any updates with this?

900miles commented 1 week ago

Yes; ecapa was trained with a shorter context window (subpart 1) and has been evaluated on speaker verification. We are working on evaluating it for diarization now. That is the context for the CHILDES dataset project, as well as the generating synthetic dataset project I am working on currently. Once we have some good datasets, I can use those to evaluate the model for diarization. We can also use those datasets for subpart 2 afterwards.