jhlusko / ChimeraNet

Unofficial implementation of music separation model by Luo et.al.

ChimeraNet

An implementation of music separation model by Luo et.al.

Getting started

Sample separation task with pretrained model

0. Prepare .wav files to separate.

1. Install library pip install git+https://github.com/leichtrhino/ChimeraNet

2. Download pretrained model.

3. Download sample script.

4. Run script

python chimeranet-separate.py -i ${input_dir}/*.wav \
    -m model.hdf5 \
    --replace-top-directory ${output_dir}

Output in nutshell

• the format of filename is ${input_file}_{embd,mask}_ch[12].wav.
• embd and mask indicates that it was inferred from deep clustering and mask respectively.
• ch1 and ch2 are voice and music channel respectively.

Train and separation examples

See Example section on ChimeraNet documentation.

Install

Requirements

• keras
• one of keras' backends (i.e. TensorFlow, CNTK, Theano)
• sklearn
• librosa
• soundfile

Instructions

1. Run pip install git+https://github.com/leichtrhino/ChimeraNet or any python package installer. (Currently, ChimeraNet is not in PyPI.)
2. Install keras' backend if the environment does not have any. Install tensorflow if unsure.

See also

• Luo et.al.
• Hershey et.al.
• Original ChimeraNet site