v-vector-tf

Tensorflow and kaldi implementation of our Interspeech2019 paper VAE-based regularization for deep speaker embedding

note: the repo is not the final release, I will clean up our experiemental code and update soon

Dependency

1. computer
2. Linux (centos 7)
3. conda (Python 3.6)
4. Tensorflow-gpu 1.8
5. kaldi-toolkit

Datasets and X-vector

1. VoxCeleb
2. SITW
3. CSLT_SITW

Steps

1. use kaldi to extract x-vector from uttrance and get xvector.ark files
2. covert the kaldi xvector.ark files to numpy binary data format (xvector.ark -> xvector.npz)
3. use tensorflow to train a VAE model, and get the V-vectors
4. use kaldi recipes to calculate EER (equal error rate)

Usage

1. install kaldi (note: if you are one of CSLT members, you can referanceDr. tzy's Kaldi or CSLT Kaldi)

2. create a conda environment and install the necessary Python package

# for example
conda create -n tf python=3.6
conda activate tf
pip install -r requirements.txt

3. git clone the code and modify the path.sh, make sure that path.sh contains your kaldi path

git clone https://github.com/zyzisyz/v-vector-tf.git

# edit path.sh
vim path.sh
# export KALDI_ROOT=${replace it by your kaldi root path}

4. calculate baseline EER

bash baseline.sh

5. Train a model

# first of all, activate the conda Python environment
conda activate tf
# you can edit train.sh to change VAE model's config
bash train.sh

6. Use kaldi-toolkit to train the backend scoring model and calculate EER

bash eval.sh

Our result

SITW Dev. Core

	Cosine	PCA	PLDA	L-PLDA	P-PLDA
x-vector	15.67	16.17	9.09	3.12	4.16
a-vector	16.10	16.48	11.21	4.24	5.01
v-vector	10.32	9.94	3.62	3.54	4.31
c-vector	9.05	8.55	3.50	3.31	3.85

Read the paper for more detail

About

Licensed under the Apache License, Version 2.0, Copyright zyzisyz

Repo Author

Yang Zhang (zyziszy@foxmail.com)

Contributors

• @Lilt
• @fatejessie
• @xDarkLemon
• @AlanXiuxiu
• @Z.K.