A Wavenet For Speech Denoising

A neural network for end-to-end speech denoising, as described in: "A Wavenet For Speech Denoising"

Listen to denoised samples under varying noise conditions and SNRs here

Installation

It is recommended to use a virtual environment

1. git clone https://github.com/drethage/speech-denoising-wavenet.git
2. pip install -r requirements.txt
3. Install pygpu

Currently the project requires Keras 1.2 and Theano 0.9.0, the large dilations present in the architecture are not supported by the current version of Tensorflow (1.2.0)

Usage

A pre-trained model (best-performing model described in the paper) can be found in sessions/001/models and is ready to be used out-of-the-box. The parameterization of this model is specified in sessions/001/config.json

Download the dataset as described below

Denoising:

Example: THEANO_FLAGS=optimizer=fast_compile,device=gpu python main.py --mode inference --config sessions/001/config.json --noisy_input_path data/NSDTSEA/noisy_testset_wav --clean_input_path data/NSDTSEA/clean_testset_wav

Speedup

To achieve faster denoising, one can increase the target-field length by use of the optional --target_field_length argument. This defines the amount of samples that are denoised in a single forward propagation, saving redundant calculations. In the following example, it is increased 10x that of when the model was trained, the batch_size is reduced to 4.

Faster Example: THEANO_FLAGS=device=gpu python main.py --mode inference --target_field_length 16001 --batch_size 4 --config sessions/001/config.json --noisy_input_path data/NSDTSEA/noisy_testset_wav --clean_input_path data/NSDTSEA/clean_testset_wav

Training:

THEANO_FLAGS=device=gpu python main.py --mode training --config config.json

Configuration

A detailed description of all configurable parameters can be found in config.md

Optional command-line arguments:

Additional arguments during inference:

Dataset

The "Noisy speech database for training speech enhancement algorithms and TTS models" (NSDTSEA) is used for training the model. It is provided by the University of Edinburgh, School of Informatics, Centre for Speech Technology Research (CSTR).

1. Download here
2. Extract to data/NSDTSEA