Vggvox-TensorFlow

This model is for speaker identification.

This model could be trained and test pre-trained on the VoxCeleb(1) datasets as described in the following paper:

[1] A. Nagrani*, J. S. Chung*, A. Zisserman, VoxCeleb: a large-scale speaker identification dataset, 
INTERSPEECH, 2017

Demo

http://chuyuan.vipgz1.idcfengye.com/

Dependencies

[1] tensorflow-gpu=1.11.0

[2] librosa=0.6.3

[3] scipy=1.1.0

Platform

[1] Ubuntu 16.04

[2] Tesla P40

[3] Python 3.5

Train

Before training and testing, you must prepare the dataset and configure the environment well.

The full dataset can be freely downloaded from VoxCeleb.

After get the full dataset, and you can see [utils/vox1_split_backup.txt] in this repo so just prepare a gpu environment, get start training!

Training:

$ python3 train.py train --voxceleb_wav_dir [wav_dir] --vox_split_txt_file [split_file] --batch_size [bs] --lr [lr] --ckpt_save_dir [save_dir]

The args:

1. --voxceleb_wav_dir: After you get full dataset, you will find all data is in a wav dir. Record the dir path.
2. --vox_split_txt_file: You can find this file in [utils/vox1_split_backup.txt], and you can also find it in VoxCeleb.
3. --batch_size: Batch size.
4. --lr: Learning rate. The default optimizer is Adam.
5. --ckpt_save_dir: Where you want to save the ckpt files. Default max ckpt files is 3.

Test

Before test, you must have the pre-trained ckpt file.

Test:

$ python3 train.py test --voxceleb_wav_dir [wav_dir] --vox_split_txt_file [split_file] --batch_size [bs] --ckpt_restore_file [ckpt_file]

The args:

1. --voxceleb_wav_dir: After you get full dataset, you will find all data is in a wav dir. Record the dir path.
2. --vox_split_txt_file: You can find this file in [utils/vox1_split_backup.txt], and you can also find it in VoxCeleb.
3. --batch_size: Batch size.
4. --ckpt_restore_file: Pre-trained model.

Example

Training:

$ python3 train.py train --voxceleb_wav_dir '/data/ChuyuanXiong/up/wav/' --vox_split_txt_file 'utils/vox1_split_backup.txt' --batch_size 32 --lr 0.001 --ckpt_save_dir '/data/ChuyuanXiong/backup/speaker_real318_ckpt' 

Testing:

$ python3 train.py test --voxceleb_wav_dir '/data/ChuyuanXiong/up/wav/' --vox_split_txt_file 'utils/vox1_split_backup.txt' --batch_size 32 --ckpt_restore_file '/data/ChuyuanXiong/backup/triplet_backup2/Speaker_vox_iter_51500.ckpt' --random_seed 100

Dataset

Set	# POIs	Utterances
Train	1251	145265
Test	1251	8251
Total	1251	153516

Dataset	My Result
voxceleb1	acc=0.7, eer=0.08

Models

Model, password: 8gy0 trained by me, and this could be used in this repo. The accuracy is about 70%, this can be further optimized. Verification eer=8%.

Model trained by the author, this is for Matlab.

Citation

@InProceedings{Nagrani17,
  author       = "Nagrani, A. and Chung, J.~S. and Zisserman, A.",
  title        = "VoxCeleb: a large-scale speaker identification dataset",
  booktitle    = "INTERSPEECH",
  year         = "2017",
}