Speech Embeddings

Introduction

This is a PyTorch implementation of a self-attentive speaker embedding.

Dataset

VoxCeleb1 contains over 100,000 utterances for 1,251 celebrities, extracted from videos uploaded to YouTube.

	dev	test
# of speakers	1,211	40
# of utterances	148,642	4,874

Download following files into "data" folder:

• vox1_dev_wav_partaa
• vox1_dev_wav_partab
• vox1_dev_wav_partac
• vox1_dev_wav_partad
• vox1_test_wav.zip

Then concatenate the files using the command:

$ cat vox1_dev* > vox1_dev_wav.zip

Dependency

• Python 3.5.2
• PyTorch 1.0.0

Usage

Data Pre-processing

Extract vox1_dev_wav.zip & vox1_test_wav.zip:

$ python extract.py

Split dev set to train and valid samples:

$ python pre_process.py

Train

$ python train.py

If you want to visualize during training, run in your terminal:

$ tensorboard --logdir runs

Performance

Model	Margin-s	Margin-m	Test(%)	Inference speed
1	10.0	0.2	88.48%	18.18 ms

Demo

Visualize speaker embeddings from test set:

$ python visualize.py

Embeddings

Theta j Distribution