Reproduce “Group Communication with Context Codec for Lightweight Source Separation” on OpenResearch

About this paper

Yi Luo, Cong Han, Nima Mesgarani

Columbia University researchers developed the Group Communication with Context Codec (GC3) design to create lightweight speech separation models, significantly reducing model size (down to 2.5% of baselines) and computational complexity (down to 17.6% MAC operations) across diverse neural network architectures while often maintaining or improving separation performance.

Key insights

• Sharing a smaller separation module across low-dimensional feature groups via GroupComm effectively minimizes model parameters and can capture inter-group dependencies.
• Compressing the temporal sequence of features with a context encoder and decoder substantially reduces computational cost, enabling faster processing and lower memory usage without significant performance loss.
• The Transform-average-concatenate (TAC) architecture proved to be the most efficient and performant design for the inter-group dependency module within GroupComm.
• Implementing a 50% overlap between feature groups further enhances separation performance while maintaining the efficiency benefits of the GC3 design.

Summary by alphaXiv.