Author: Jiachuan Deng

Goal

On-going project experimental codes for building a real time one-shot deep learning speaker identification system.

File Structure

Speaker-Identification
|-- ReadMe.md
|-- data
|   |-- voxData (raw data downloaded from http://www.robots.ox.ac.uk/~vgg/data/voxceleb/vox1.html)
|   |-- voxVadData (raw data after vad)
|   
`-- src
    |-- dataprocessing
    |    |-- config.ini (config file)
    |    |-- run_dataprocessing.py (exe code)
    |    |
    |    `-- vad.py (vad toolkit)
    |
    |-- model
    |     |
    |     `-- models.py (definemodels)
    |-- utils
    |     |-- autioDataset.py (Data set defined in pytorch dataset style)
    |     |
    |     `-- extractFeature.py (extract MFCC feature from raw audio)
    |
    `-- trainModel.py (executable script to train model)
  

Run code

0. data processing

under dataprocessing directory, do:

python3 run_dataprocessing.py

Make sure the paths in config.ini are consistent with your local files.

The script will result in several audio chunks saved under /data/voxVadData, and a train_file_path.txt,val_file_path.txt stored under /dataprocessing.

1. train model

under Speaker-Identification directory, do: i. for train classifier encoder model:

python3 trainModel.py -m 0 

ii. for train siamesenet model:

python3 trainModel.py -m 1