search

mycrazycracy / tf-kaldi-speaker

Neural speaker recognition/verification system based on Kaldi and Tensorflow
Apache License 2.0
32 stars 16 forks source link

About "GE2E loss" #1

Closed PES2g closed 5 years ago

PES2g commented 5 years ago

I saw code for GE2E loss in the code which have already been commented. Do you guys try GE2E loss in the experiment ? If yes, what is the performance of GE2E loss ?

mycrazycracy commented 5 years ago

Hi, the end2end loss performs poorly in my experiment (compared with softmax or amsoftmax). I suspect that this loss is not a good option when the dataset is not too large (#number of speakers is small). So I just commented on the related code.

00001101-xt commented 5 years ago

I saw code for GE2E loss in the code which have already been commented. Do you guys try GE2E loss in the experiment ? If yes, what is the performance of GE2E loss ?

End-to-End requires large training set(greater than 10k individual identities at least I think) which isn't so easy to obtain. So no, the GE2E loss won't work too well with a small dataset.