One hand-crafted topological feature makes speech neural nets far more noise-robust

bravo_abad · x · 2026-09-21

A new paper by Feng et al. shows that converting a speech signal into a geometric object via time-delay embedding and using persistent homology to measure the persistence of its dominant structures yields a single hand-crafted topological feature that markedly improves a speech network's noise robustness — challenging the assumption that deep nets always learn such features themselves. The author also runs the Substack newsletter Discovery at Scale, curating actionable AI-for-Science ideas weekly.

Related event: Handcrafted Topological Feature Boosts Speech Network Noise Robustness(2 posts)→

Original post →

More from Research

Research channel →