Parallel Time-Band Mixing with Learned Observation-Adding for Robust ASR Front-Ends
2026-08-31 • Sound
SoundArtificial Intelligence
AI summaryⓘ
The authors designed a new speech enhancement method that works faster by processing time and frequency parts of audio in parallel instead of one after another. Their approach, called Parallel Time-Band Mixer (PTBM), combines mixing within frequency bands over time and attention across bands at each moment simultaneously. They also added a technique to reduce errors sensitive to speech recognition without extra tuning. Tests show their method improves automatic speech recognition accuracy while keeping the model small and efficient.
speech enhancementautomatic speech recognition (ASR)Parallel Time-Band Mixer (PTBM)cross-band attentiontemporal mixingmask-and-residual reconstructionObservation-Addingword error rateparallel processingmodel efficiency
Authors
Xingyu Shen, Runze Wang, Wei-Ping Zhu, Benoit Champagne
Abstract
Speech enhancement is often used as a front-end for robust ASR, yet recurrent temporal and cross-band modules introduce sequential dependencies that reduce parallel efficiency. In this paper, we present a sequence-parallel band-split enhancement front-end built on a Parallel Time-Band Mixer (PTBM) block that eliminates within-block recurrent unrolling. PTBM integrates intra-band temporal mixing and per-frame cross-band attention within a unified parallel architecture, enabling efficient contextual modeling across both time and frequency dimensions. The system retains the mask-plus-residual reconstruction interface and introduces learned Observation-Adding (LOA) to suppress ASR-sensitive artifacts without development-set tuning. Experiments on DNS Challenge and CHiME-4 with frozen Whisper back-ends show that the proposed front-end consistently reduces word error rate relative to recurrent band-split baselines while requiring only 0.96 M parameters and 0.58 GMAC/s for the front-end network.