nnnoiseless

nnnoiseless is a rust crate for suppressing audio noise. It is a rust port of the RNNoise C library, and is based on a recurrent neural network.

While nnnoiseless is meant to be used as a library, a simple command-line tool is provided as an example. It operates on WAV files or RAW PCM files. Run

cargo install nnnoiseless

to install it (you might need to install rust first). Once nnnoiseless is installed, you can run it like

nnnoiseless input.wav output.wav

or, for more advanced usage, try

nnnoiseless --help

Safety

Except for the C API described below, nnnoiseless is mostly written in safe rust. It currently uses unsafe in two places, to cast arrays of f32s to arrays of Complex<f32>s with half the length; this delivers a small but measurable performance improvement. If a future version of RustFFT has built-in support for real-only FFTs, this unsafe code will be removed.

C API

It is possible to install nnnoiseless as a library usable from C, with an RNNoise-compatible header.

$ cargo install cargo-c
$ mkdir staging-nnnoiseless
$ cargo cinstall --destdir staging-nnnoiseless
$ sudo cp -a staging-nnnoiseless/* /

Custom models

nnnoiseless is based on a neural network. There's one built in, but you can also swap out the built-in network for your own. (This might be useful, for example, if you have a particular kind of noise that you want to filter out and nnnoiseless's built-in network doesn't do a good enough job.)

Loading a nnnoiseless network

Let's suppose that you've already trained (or downloaded from somewhere) your neural network weights, and that they are in the file weights.rnn. You can use these weights for the nnnoiseless binary by passing in the --model option:

nnnoiseless --model=weights.rnn input.wav output.wav

On the other hand, if you're using nnnoiseless as a library, you can load your neural network weights using RnnModel::from_bytes or RnnModel::from_static_bytes.

Converting an RNNoise network

Some people have already made their own neural network weights for RNNoise (for example, here). These weights can be used in nnnoiseless also, but you'll need to first convert them from the (text-based) RNNoise format to the (binary) nnnoiseless format. There is a script in the train directory that can do this for you: just run

python train/convert_rnnoise.py input_file.txt output_file.rnn

Training your own weights

This is a little involved, but at least it's documented now. See train/README.md for more information.