This is a small repo to make it easier to use SE-FFTNet to clean up audio samples. It wraps the model in Docker, takes in noisy WAV files and outputs cleaned up WAVs.

Requirements:

• Linux (currently, until CUDA for WSL2 is released)
• Docker & Nvidia-docker (tested on 19.03.6)

The container follows the requirements of SEGAN. Ideally this would use a Tensorflow Docker image but there are some issues in getting this to run. Instead, it is built on an NVidia CUDA image, specificially this one.

We also need to use the weights provides by SEGAN - you can find these here. These should be unzipped, with segan_v1.1 being placed in a directory called weights. This should be in the directory you build the container from, with the directory structure looking something like this:

SEGAN-Docker/
├── build_gpu.dockerfile
├── repositories/
│   ├── ...
├── weights/
│   ├── segan_v1.1
│   |   ├── SEGAN-41700.data-00000-of-00001
│   |   ├── SEGAN-41700.index
│   |   ├── SEGAN-41700.meta
├── segan/
│   ├── ...

Create two local directories - one for input, one for output. You will need to mount these when you run the container. It should look something like this:

output/
input/
├── input_file_1.wav
├── input_file_2.wav
├── input_file_3.wav

If you are happy to mount from your clone of this repo, for example, you could use /path/to/repo/segan/input/ and /path/to/repo/segan/output/, and your SEGAN-Docker directory may look like this:

SEGAN-Docker/
├── build_gpu.dockerfile
├── output/
├── input/
│   ├── input_file_1.wav
│   ├── input_file_2.wav
│   ├── input_file_3.wav
├── repositories/
│   ├── ...
├── weights/
│   ├── segan_v1.1
│   |   ├── SEGAN-41700.data-00000-of-00001
│   |   ├── SEGAN-41700.index
│   |   ├── SEGAN-41700.meta
├── segan/
│   ├── ...

Note: Ensure that output does not contain a file with the same name as whatever you pass as input, or SEGAN may fail to save the file.

You can build with:

docker build -t segan:<version number> -f build_gpu.dockerfile .

Put the files you wish to convert in input, ensuring they are sampled at 16 KHz. Then run the container, mounting it as below and giving the name of a file for processing <input file name> in the input directory:

docker container run --runtime=nvidia -v <local input path>:/segan/input/ -v <local output path>:/segan/test_clean_results/ -e input_file=<input file name> segan:<version number>