Skip to content
master
Switch branches/tags
Go to file
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 

README.md

Vector-Quantized Autoregressive Predictive Coding

For an improved version of VQ-APC, please refer to this repository.

This repository contains the official implementation of Vector-Quantized Autoregressive Predictive Coding (VQ-APC).

VQ-APC is an extension of APC, which defines a self-supervised task for learning high-level speech representation from unannotated speech. For dependencies and data preprocessing, please refer to the implementation of APC. After the data are ready, here's an example command to train your own VQ-APC model:

python train_vqapc.py --rnn_num_layers 3 \
                      --rnn_hidden_size 512 \
                      --rnn_dropout 0.1 \
                      --rnn_residual \
                      --codebook_size 128 \
                      --code_dim 512 \
                      --gumbel_temperature 0.5 \
                      --apply_VQ 0 0 1 \
                      --optimizer adam \
                      --batch_size 32  \
                      --learning_rate 0.0001 \
                      --epochs 10 \
                      --n_future 5 \
                      --librispeech_home ./librispeech_data/preprocessed \
                      --train_partition train-clean-360 \
                      --train_sampling 1. \
                      --val_partition dev-clean \
                      --val_sampling 1. \
                      --exp_name my_exp \
                      --store_path ./logs

Argument descriptions are available in train_vqapc.py.

TODOs

  • Add scripts that get the learned codebook(s) (essentially the parameters of the nn.Linear layer used to implement the VQ layers)
  • Add scripts that visualize the code-phone co-occurrence (Figure 3 in the paper)

Reference

Please kindly cite our work if you find this repository useful:

@inproceedings{chung2020vqapc,
  title = {Vector-quantized autoregressive predictive coding},
  autohor = {Chung, Yu-An and Tang, Hao and Glass, James},
  booktitle = {Interspeech},
  year = {2020}
}

Contact

You can reach me out via email. Questions and feedback are welcome.

About

Vector Quantized Autoregressive Predictive Coding (VQ-APC)

Topics

Resources

Releases

No releases published

Packages

No packages published

Languages