MLASS

This project extended the LASS method to support more than two sources while maintaining feasible memory complexity.

LASS: The Original Work

LASS separates mixed sources without needing additional gradient-based optimization. It uses a VQ-VAE to embed signals into a discretized latent space and autoregressive priors to sample original sources from a joint posterior.

MLASS: My Extension

I proposed two methods to decouple memory complexity from the number of sources ($n$), significantly optimizing the original $O(k^n)$ complexity.

• 01
  Belief Propagation (BP):

  Reduces memory complexity to $O((n-1)k^3)$, making multi-source separation viable on standard hardware.

• 02
  Probabilistic Extractor (PE):

  An alternative sampling approach for source separation in latent space.

2 Sources: Original - BP - PE

3 Sources: Original - BP - PE

Quantitative Benchmarks