DaNet-Tensorflow

Tensorflow implementation of "Speaker-Independent Speech Separation with Deep Attractor Network"

Link to original paper

STILL WORK IN PROGRESS, EXPECT BUGS

Requirements

numpy / scipy

tensorflow >= 1.2

matplotlib (optional, for visualization)

h5py / fuel (optional, for certain datasets)

Usage

Prepare datasets

Currently, TIMIT and WSJ0 datasets are implemented. You can use the "toy" dataset for debugging. It just some white noise.

TIMIT dataset

Follow app/datasets/TIMIT/readme for dataset preparation.

WSJ0 dataset

Follow app/datasets/WSJ0/readme for dataset preparation.

After setting up a dataset, you may want to change DATASET_TYPE in hyperparameters.

Setup hyperparameters

This is to change batch size, learning rate, dataset type etc ...

The recommended way: using JSON file

There's a default.json file at the root directory. You make your own and change some of the values. For example you can create a JSON file with:

{
    DATASET_TYPE="timit",
    LR=1e-2,
    BATCH_SIZE=8
}

Save it as my_setup.json, now you can run the script with:

python main.py -c my_setup.json

The direct way: using command line arguments

Some commonly used hyperparameters can be overridden by CLI args.

For example, to set learning rate:

python main.py -lr=1e-2

Here's a incomplete list of them:

# set learning rate, overrides LR
-lr
--learn-rate

# set dataset to use, overrides DATASET_TYPE
-ds
--dataset

# set batch size, overrides 
-bs
--batch-size

# set

Note If you get out of memory (OOM) error from tensorflow, you can try using a lower BATCH_SIZE.

Note If you change FFT_SIZE, FFT_STRIDE, FFT_WND, SMP_RATE, you should do dataset preprocessing again.

Note If you change model architecture, the previously saved model parameter may not be compatible.

Perform experiments

Under the root directory of this repo:

train a model for 10 epoch and see accuracy, using TIMIT dataset

    python main.py -ds='timit'

train a model using your own hyperparameters

    python main.py -c my_setup.json

train a model for 100 epoch and save it

    python main.py -ne=100 -o='params.ckpt'

continue from last saved model, train 100 more epoch, save back

    python main.py -ne=100 -i='params.ckpt' -o='params.ckpt'

test the trained model on test set

    python main.py -i='params.ckpt' -m=test

draw a sample from test set, then separate it:

    $ python main.py -i='params.ckpt' -m=demo
    $ ls *.wav
    demo.wav demo_separated_1.wav demo_separated_2.wav

separate a given WAV file:

    $ python main.py -i='params.cpkt' -m=demo -if=file.wav
    $ ls *.wav
    file.wav file_separated_1.wav file_separated_2.wav

launch tensorboard and see graphs

    tensorboard --logdir=./logs/`

for more CLI arguments, do

    python main.py --help

Use custom dataset

Make a file app/datasets/my_dataset.py.
Make a subclass of app.datasets.dataset.Dataset

    @hparams.register_dataset('my_dataset')
    class MyDataset(Dataset):
        ...

You can use app/datasets/timit.py as an reference.

In app/datasets/__init__.py, add:

    import app.datasets.my_dataset

To use your dataset, set DATASET_TYPE to "my_dataset" in JSON config file

Customize model

You can make subclass of Estimator, Encoder, or Separator to tweak model.

Encoder is for getting embedding from log-magnitude spectra.
Estimator is for estimating attractor points from embedding.
Separator uses mixture spectra, mixture embedding and attractor to get separated spectra.

You can set encoder type by setting ENCODER_TYPE in hyperparameters.

You can set estimator type by setting TRAIN_ESTIMATOR_METHOD and INFER_ESTIMATOR_METHOD in hyperparameters.

You can set separator type by setting SEPARATOR_TYPE in hyperparameters.

Make sure to use @register_* decorator for your class. See code in app/modules.py for details. There are existing sub-modules.

To change overall model architecture, modify Model.build() in main.py

Limitations

Only the favorable "anchor" method for estimating attractor location during inference is implemented. During training, it's also possible to use ground truth to give attractor location.
TIMIT dataset is small, so we use same set for test and validation.
We use WSJ0 si_tr_s / si_dt_05 / si_et_05 subsets as training / validation / test set respectively. The speakers are randomly chosen and mixed at runtime.

This setup is slightly different to orignal paper.
Only single GPU training is implemented.
Doesn't work on Windows.

sa757 / DaNet-Tensorflow