sysprog21 / rnnoise

A noise suppression library based on a recurrent neural network

RNNoise

A noise suppression library based on a recurrent neural network.

Build

Prerequisite

• macOS: brew install libsndfile libsoxr sox
• Debian/Ubuntu: sudo apt install libsndfile1-dev libsoxr-dev libsox-dev

To compile, just type:

make

Sample noisy file sample.wav was included, and you can run make check to generate the processed one, clean.wav.

Test

While it is meant to be used as a library, a simple command-line tool is provided as an example. It can be used as:

examples/rnnoise_demo sample.wav output.wav

Training

Audio feature extract

Build audio feature extraction tool

make src/denoise_training

Use the tool denoise_training to get the audio feature array from speech and noise audio clip

src/denoise_training signal.raw noise.raw count > training.f32

(note the matrix size and replace 500000 87 below)

RNN model traning

Pick feature array to "training" dir and go through the training process.

cd training ; ./bin2hdf5.py ../src/training.f32 500000 87 training.h5
./rnn_train.py
./dump_rnn.py weights.hdf5 ../src/rnn_data.c ../src/rnn_data.h

Training process will generate the RNN model weight code file (default is rnn_data.c) and layer definition header file (default is rnn_data.h). They can be used to refresh the src/rnn_data.c, src/rnn_data.h and rebuild the rnnoise library and/or examples.

License

rnnoise is freely redistributable under the revised BSD license. Use of this source code is governed by a BSD-style license that can be found in the COPYING file.