ishine/danna-sep

Danna-Sep: Danna Sama no Source Separation

The winning model I used in MDX 2021 Challenge.

Installation

python setup.py install

Usage

usage: danna_sep [-h] [--outdir OUTDIR] [--fast] infile

positional arguments:
  infile           input audio file

optional arguments:
  -h, --help       show this help message and exit
  --outdir OUTDIR  output directory. Default to current working directory
  --fast           faster inference using only two of the models

Given an audio file, the program will split it into 4 stems, which are drums, bass, vocals and other, and store them in the given directory as .wav files.

When execute the first time it will download our pre-trained models (around 1 to 2 Gb) to the directory specified by the environment variable DANNA_CHECKPOINTS, which by default is ~/danna-sep-checkpoints.

This process is very time comsuming and require at least 16 Gb of RAM.

Training

Please refer to our training repo.

TODO

• convert to ONNX format
• Pack it as standalone app