search

resemble-ai / Resemblyzer

A python package to analyze and compare voices with deep learning
Apache License 2.0
2.66k stars 419 forks source link

EER of pre-trained model? #48

Closed felixkreuk closed 3 years ago

felixkreuk commented 3 years ago

Hi, thanks for this repository. I understand that the speaker embedding model is based on "Generalized End-To-End Loss for Speaker Verification" and was trained on VoxCeleb2. Could you please mention what is the EER of your pre-trained model?

Thank you

CorentinJ commented 3 years ago

Refer to the end of page 19 in my thesis.

EER evaluation was probably the shakiest part of my work, mainly because I couldn't find details of the procedure in the literature. If you're looking to have a metric to report in an academic context, I'd highly recommend you reproduce it on your own. And if you do, I'm interested in the results.

felixkreuk commented 3 years ago

I was actually interested in using your speaker encoder for generation purposes. Your demo sounds quite good, but having an objective metric is always nice :)

On 24 Feb 2021, at 14:59, Corentin Jemine notifications@github.com wrote:

 Refer to the end of page 19 in my thesis.

EER evaluation was probably the shakiest part of my work, mainly because I couldn't find details of the procedure in the literature. If you're looking to have a metric to report in an academic context, I'd highly recommend you reproduce it on your own. And if you do, I'm interested in the results.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.