RNNoise

What RNNoise is

RNNoise is a free, open-source noise suppression library based on a recurrent neural network. The project is intended for use as a library and includes a command-line demo. That demo processes mono, 16-bit raw PCM at 48 kHz and reads and writes raw files rather than WAV files. The documented build sequence uses autogen.sh, configure, and make. For CPU optimization, the project recommends AVX2 and describes AVX2 or SSE4.1 runtime CPU detection as build options. RNNoise can use its default model or load a custom model through its API from a file, filename, or memory buffer. An alternative smaller model with increased sparsity is also described. Training requires clean speech and noise data sampled at 48 kHz in 16-bit PCM format. The repository identifies the license as BSD-3-Clause and lists Linux as its platform. The README describes the GitHub repository as a convenience copy and points to GitLab for the latest source. Xiph directs people to mailing lists or IRC channels to contact the foundation and community.

Who it suits

RNNoise suits developers seeking a noise suppression library for Linux or a command-line demo. Training a model requires suitable clean speech and noise data in the specified format.

Limits

The command-line demo is limited to mono, 16-bit raw PCM at 48 kHz and uses raw files rather than WAV. Model training requires clean speech and noise data in 48 kHz, 16-bit PCM format.

Getting started

1. Open the RNNoise project on GitLab, identified as the location of the latest source.
2. Build with the autotools sequence described in the README: autogen.sh, configure, and make.
3. Use the library in your software or try the included command-line demo.
4. For the demo, prepare mono, 16-bit raw PCM audio sampled at 48 kHz.
5. Use the default model or load a custom model through the API.

Verdict

RNNoise provides a library and demo with options for default or custom models. The demo's raw-audio format and the model-training data requirements are worth checking first.