GANSASSTF

TensorFlow implementation of "GAN Single Audio Source Separation"

WORK IN PROGRESS

The model currently incomplete, having performance issues, and doesn't perform the task well ... yet.

Dependencies

[python]

• six
• numpy, scipy, nltk
• tensorflow 1.12+

[TIMIT dataset]

• install "sndfile-convert" utility sudo apt-get install sndfile-programs

[optional]

• install warp-ctc for better GPU CTC performance git clone https://github/baidu-research/warp-ctc then follow instructions inside to install

Prepare datasets

toy dataset

There is a "toy dataset" for debugging purposes. It's just pure generated white noise.

In app/hparams.py file, set DATASET_TYPE = 'toy' to use this dataset.

TIMIT

• Make sure to install sndfile-tools first (for sndfile-convert utility)
• Download the TIMIT.zip file under app/datasets/TIMIT/.
• Under repo directory app/datasets/TIMIT/, run install.sh. You only need to run it for once.
• If you change hyperparameter FFT_SIZE or CHARSET_SIZE, you need to re-run install script.
• In app/hparams.py file, set DATASET = 'timit' to use this dataset

Quick Instructions

To quickly train the model, modify app/hparams.py file to setup hyperparameter, then run python main.py -o saves/my_model.cpkt.

This will will save model parameters in saves/my_model.cpkt

To continue training, run python main.py -i saves/my_model.cpkt

To check model accuracy on test set, do python main.py -i <YOUR_MODEL_PARAMETERS> -m=test

To quickly use the model to separate a WAV file, do python main.py -i <YOUR_MODEL_PARAMETERS> -m=demo -if=your_audio.wav.

For more CLI arguments help, do python main.py --help

Code structure

main.py defines global model architecture, training / testing, and basic CLI funcitonality

app/datasets contains code to preprocess / access datasets.

app/hparams.py defines hyperparameters for quick finetune.

NOTE ASR module is currently buggy.

This defines common hyperparameters such as learn rate, batch size. As well as submodule types to use.

app/modules.py defines architecture for sub-modules that are used in the model, such as separator, discriminator. You can add your own sub-modules here. For example:

@register_separator('my_separator')
class MySeparator(Separator):
    ...

Then change hparams.py:

SEPARATOR_TYPE = 'my_separator'

app/ops.py defines some commonly used Tensorflow Ops.

app/ozers.py defines optimizers.

app/utils.py defines commonly used subroutines.

saves/ directory intends to store model parameters.

Run clear_saves.sh to remove everything there. You can put them elsewhere though.

logs/ directory stores TF summaries.

Start TensorBoard via tensorboard --logdir=logs/ to visualize them.

asr_logs/ directory stores TF summaries for speech recgonizer.

Coding style

This codebase mostly follow PEP8, but there are a few exceptions.

For symbolics:

• s_var means a placeholder or symbolic element in a graph.
• sS_var means it's also sparse.
• v_var means a tensor variable with value, typically parameter.