Unsupervised Blind Source Separation with Variational Auto-Encoders
Julian Neri, Roland Badeau, Philippe Depalle
This work was presented at EUSIPCO 2021. The paper, video presentation and slides are available here:
VAE Model
Results
MNIST – 2 Ground Truth Sources – 2 Model Sources
MNIST – 2 Ground Truth Sources – 3 Model Sources
Visualization
Inferred latent sources from the MNIST dataset were mapped to two dimensions with multi-dimensional scaling (MDS). MDS seeks a 2D representation that respects the distances of the original D_z-dimensional latent sources.
The dense collection of points at (0,0) in the right two pictures correspond to trimmed sources that generate black images (every pixel being approximately zero, as seen in the above examples for VAEM).
K=2
K=3
K=4
Audio Examples
Example 1 – Two Violins with Vibrato
Example 2 – Acoustic Guitar and Alto Sax Growl
Example 3 – Bassoon and Violin with Vibrato
Example 4 – Snare Drum Roll and Cello
Source Code
PyTorch source code and datasets is available on my GitHub.
Citation
J. Neri, R. Badeau, P. Depalle, “Unsupervised Blind Source Separation with Variational Auto-Encoders”, 29th European Signal Processing Conference (EUSIPCO), Dublin, Ireland, August 2021.
BibTex
@inproceedings{Neri:2021:VAE:Unsupervised,
author = {Julian Neri and Philippe Depalle and Roland Badeau},
title = {Unsupervised Blind Source Separation with Variational Auto-Encoders},
booktitle = {29th European Signal Processing Confernce (EUSIPCO)},
address = {Dublin, Ireland},
pages = {311-315},
month = {August},
year = {2021}}