MSc Statistics Project - Audio Source Separation

The Audio Source Separation task aims to separate an audio mixture of different sources. In this project, we focus on a Bayesian approach using Deep Generative Priors. We try to separate Beethoven and Bach Sonata of Piano and Violin.

Link to the Thesis

NCSN Results

Here are some extracts generated by the Noise Conditionned Score Network (NCSN) model. This model is trained to generate 2 seconds spectrograms which are then converted to audio signal.

Piano

Violin

BASIS Separation

Beethoven Sonata 1: 1st minute. The generative model used is the NCSN. The original sources and mixture are obtained with the same method as the estimated sources, i.e., inversion of the melspectrograms.

Originals Sources

Mixture

Estimated Sources

After applying a Single-channel Wiener Filter