Spleeter Baseline
U-Net based audio source separation by Deezer
Experiment with different loss functions to compare their performance. Different spectro losses: mel-spectrogram and magnitude. Different distance costs: L1 and MSE.
logs contains training loss log with tensorboard
data.py dataloader, experimented with both original baseline Spleeter dataloader and Open-Unmix dataloader dataloader
display_mask.ipynb contains what model outputs (a ratio mask) and spectrogram comparison between ground truth audio source (vocal) with separated audio source (mixture * ratio mask)
run.py model training code
splitter.py inferencing code
