Source separation neural network

This is the implementation of the source separation neural network which employs the U-net structure. The implementation refers to Singing Voice Separation with Deep U-Net Convolutional Networks.

Model architecture

The encoder and decoder blocks of the U-net each contains 6 convolutional blocks. The model takes the STFT magnitude spectrogram of the input signal and outputs masked STFT spectrogram.

Folder structure

source_separation
+--README.md
+--mask_data
|  +--mixtures
|     +--train
|     +--val
|     +--test
|  +--targets
|     +--train
|     +--val
|     +--test
+--model
+--pickle_data
|  +--train
|  +--val
|  +--test
+--src_formatted
+--test_result

Installation

To run the code, python, pytorch, torchaudio, numpy, and librosa are required.

How to run

1. Have the data in the mask_data folder as the structure above. Every sample in training and validation set must of equal length for batch processing.
2. Run serialize.py to obtain the pickle data.
3. Run mask_main.py to execute the training and inference.

Authors

Diep Luong