A Multichannel Band-split Recurrent Architecture for Real-time Speaker-conditioned Speech Enhancement
Year: 2026
Type: conference-paper
Abstract: This paper presents MBSRNN, a real-time multichannel speech enhancement and speaker extraction system developed for the CHiME-9 ECHI challenge.MBSRNN efficiently leverages spatial and phase information across microphones.Unlike traditional methods, bandsplit processing to the enrollment speech is applied prior to encoding, which allows us to capture fine-grained speaker characteristics at the subband level.This speaker information is integrated into the network via Feature-wise Linear Modulation (FiLM) interleaved with LSTM blocks.To ensure online-streaming capability for hearing assistive devices, we employ a chunk-based inference strategy with a strict maximum algorithmic latency of 20 ms.Experimental results on the CHiME-9 development set demonstrate that MBSRNN effectively extracts target speech in highly noisy environments while maintaining robust real-time performance.
Authors: Zeen Jhou, Rong Chao, You-Jin Li, Yu Tsao
Institutions: National Taiwan University, Academia Sinica
Language: English
Topic: Speech and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W7164667751