LaCo-SENet hits causal SOTA speech quality at 12.5ms latency

Streaming speech enhancement has always faced a trade-off: wait for future audio to improve quality, or process instantly and accept degraded output. LaCo-SENet, from researchers Yunsik Kim and Yoonyoung Chung, breaks this binary. It introduces asymmetric temporal padding, which redistributes past and future context within convolutions, plus a dual-buffer streaming mechanism that combines state buffers for past context with lookahead buffers for future context at both input and feature levels. A single training-time hyperparameter lets developers configure the exact latency point from fully causal 12.5ms up to 75ms, all from the same 1.37M-parameter backbone.

The team validated LaCo-SENet on VoiceBank+DEMAND, a standard speech enhancement benchmark. The resulting model family scores PESQ 3.35 at 12.5ms (fully causal) to 3.43 at 75ms. Critically, the 12.5ms causal model beats the prior causal state-of-the-art (PESQ 3.27 at 46.5ms) while running at nearly a quarter of the latency. Selective state updates prevent future-frame leakage, ensuring training-inference consistency. Accepted at Interspeech 2026, this work gives real-time voice applications a practical knob: dial in sub-15ms latency for live calls, or relax to 75ms for higher quality in recordings, without swapping models.