An SVD view of Conv-TasNet: signal-processing structure, spectral dynamics, and low-rank compression

When: Wednesday, August 5, 2026, 3:00 PM - 4:00 AM
Where: Engineering G2, Mile End

Abstract: Conv-TasNet is often presented as a neural time-domain separator, but its design is deeply rooted in acoustic signal processing: a learned analysis-synthesis filter bank replaces the STFT, masking is retained as the separation mechanism, and temporal convolutions implement structured finite-impulse-response filtering. This work studies Conv-TasNet through singular value decomposition (SVD). For depthwise temporal convolutions, we view each filter as a Toeplitz operator, linking its singular spectrum to the filter frequency response through Szegő-type theory. For pointwise 1 × 1 convolutions, we analyze the evolution of singular values and singular subspaces during training, showing rapid early spectral concentration and progressive stabilization. We then summarize a low-rank factorization scheme for the pointwise layers. On Libri2Mix, truncation near the effective rank gives the strongest compression-performance trade-off, while more aggressive truncation degrades gracefully. These results position Conv-TasNet as both an interpretable signal-processing architecture and a compressible low-rank model. This work was conducted as part of Jingwei's postdoctoral research at Queen Mary. The corresponding paper has been accepted for publication at the International Workshop on Acoustic Signal Enhancement (IWAENC) and will be presented in Cremona, Italy, in September 2026.

Bio: Jingwei Liu is a Postdoctoral Research Associate at the Centre for Digital Music (C4DM) and the Centre for Fundamentals of AI and Computational Theory in the School of EECS at QMUL. She works under the EPSRC-funded grant, Artificial Neuroscience: metrology and engineering for Deep Learning using Linear Algebra, supervised by the grant PI, Prof. Mark Sandler. Before joining Queen Mary in 2025, Jingwei completed her Ph.D. in Computer Music at the University of California, San Diego. Jingwei's research lies at the intersection of artificial intelligence, audio and music technology, signal processing, and cognitive neuroscience. Her current work investigates the internal structure and training dynamics of audio neural networks using spectral analysis, singular value decomposition, and low-rank modelling, with the aim of developing more interpretable, efficient, and compact AI systems. Her broader research interests include expressive music generation, real-time human–AI musical interaction, predictive processing, active inference, computational creativity, and neuro-inspired machine learning.