WaveUnet Implementation for Saraga Dataset (Indian Carnatic Music)

Actual Network: https://github.com/f90/Wave-U-Net-Pytorch. We have used this as a baseline and restructured it to use the appropriate dataset.

Saraga Carnatic Dataset:

It has five stems: Mixed, Vocal, Violin, Mrindangam right, and Mrindangam Left. Converting Mridangam left and right into a single audio file (mridangam) Expecting Four stem outputs, namely: Vocal, violin, mridangam and others

Version 1: 4 Sources (Leakage)

The dataset is trained to extract stems: Mrindangam Left, Mrindangam Right, Vocal and Violin.

Results:

Metrics	With Bleeding Effects
SDR	-0.19096690404060424

Version 2: 3 Sources (Leakage)

The dataset is trained to extract Three stems: Mridangam, Vocal and Violin. This is with some minor changes in the code in data loading. The mridangal left and right are added together. Where ever there is a secondary vocal, it is added to the primary vocal. Ghatam files are removed.

Results:

Metrics	With Bleeding Effects
SDR	1.166956417870889

Version 3: 3 Sources (Leakage Removed)

The dataset is trained to extract stems: Mridangam (left+right), Vocal(/s) and Violin, Bleeding of the sources is reduced considerably to achieve higher performance.

Results:

Metrics	With Bleed	Without Bleed
SDR	1.166956417870889

Standard MUSDB18HQ

MUSDB18HQ is internally artificially bled to evaluate the performance of the effect of leakage. The actual source is maintained dominant and other sources are bled to it with a volume reduction of 10dB. Trained on PyTorch: Wave-U-Net network.

Results:

Metrics	Actual	With Bleed	After Bleed Removal
SDR	2.309013108265547	0.9656615837211856	1.729928257040701