kjw11 / tf-kaldi-speaker

Neural speaker recognition/verification system based on Kaldi and Tensorflow

TF-kaldi-speaker

This code is forked from entn-at/tf-kaldi-speaker. It is a speaker verification system based on Kaldi and TensorFlow. More detail please refer: entn-at/tf-kaldi-speaker.

Features

This version has two features compared with the original branch:

Resnet-34 Topology

This is a famouse Resnet topology, the blocks are: [3/32, 3/32], [3/64, 3/64], [3/128, 3/128], [3/256, 3/256], and the number of blocks is: [3, 4, 6, 3]

The code is /model/resnet.py.

SITW Recipe

A SITW recipe is added in egs/sitw, which is largely based on SITW offical x-vector recipe.

There are 8 exprimental settings in the SITW recipe, with different network topologies, pooling methods and loss functions. See ./egs/sitw/v1/nnet_conf. Note that the training and test data are the same as SITW offical recipe.

Some of the experimental results are shown below:

Topoloy	Pooling	Loss func	EER(%)
TDNN	Statistic Pooling	Softmax	2.43
TDNN	Attention Pooling	AAM-Softmax	2.49
TDNN	Statistic Pooling	Softmax	2.41
TDNN	Attention Pooling	AAM-Softmax	2.57
Resnet-34	Statistic Pooling	Softmax	2.41
Resnet-34	Attention Pooling	AAM-Softmax	1.96
Resnet-34	Statistic Pooling	Softmax	2.16
Resnet-34	Attention Pooling	AAM-Softmax	2.30

Where "AAM" means additive angular margin.