Spleeter

Implementation of Spleeter by PyTorch

Dependencies

1. Python2/3
2. pip install -r requirements.txt

Usage

Training steps

1. python preprocess.py

• Need to fill two parameters train_dataset and train_manifest

​ I. train_dataset: the PATH of training set The directory structure is recommonded to be:

├── Dataset
|    ├──song1
|    |	 ├── mixture.wav
|    |   ├── vocals.wav
|    |   ├── instrumental.wav
|    ├── song2
|    |   ├── mixture.wav
|    |   ├── vocals.wav
|    |   ├── instrumental.wav

​ This means one folder only contains one songs, including three audios (mixture, vocal and background music)

​ II. train_manifest: contains song information, utilized by train.py.

2. python train.py

You can use params to adjust training parameters.

Notice:

train_manifest: the PATH of training manifest

load_model: three optional variables, including:

I.  “tensorflow”：trains with the pre-trained model trained by tensorflow
II.  “pytorch”：trains with the pre-trained model trained by PyTorch
III. None：trains without a pre-trained model

Prediction steps

Notice

I. If you use the model in tensorflow, need 2stems, model.py, util.py, separator.py

II. If you use the model in PyTorch, need final_model/net_vocal.pth, final_model/net_instru.pth, model.py, util.py, separator.py

Separation model is encapsulated in class Separator of separator.py：

1. from separator import Separator
2. sep = Separator(params(optional))
3. sep.separate(input_wav_path(path of the target audio，MUST)， output_dir(output path of audios，optional))

Reference

1.[Music Source Separation tool with pre-trained models / ISMIR2019 extended abstract] (http://archives.ismir.net/ismir2019/latebreaking/000036.pdf)