Contextual biasing improves speech recognition for unsegmented languages
Boundary-Free Contextual Biasing: Depth-Adaptive Gating and Reading-Space Matching for Unsegmented Languages
Computation and LanguageSound
Summary
Speech recognition systems help convert spoken words into text but often rely on knowing where words start and end. This is hard for languages like Chinese and Japanese, which do not clearly separate words. The authors propose a new way to bias these systems toward expected words without depending on word boundaries, using special character-level techniques instead. Their method improves the system’s ability to correctly recognize rare or tricky words without additional training. They also created the first public benchmark dataset for Japanese speech recognition with this contextual biasing.
What this means in practice
- •For speech recognition developers: Implement contextual biasing in systems for Chinese and Japanese to better recognize rare words without retraining models.
- •For voice assistant teams: Improve accuracy of voice assistants in East Asian languages by integrating boundary-free biasing to handle unsegmented speech input.$Commercial implications: Enables selling more accurate voice assistant products tailored to Japanese and Chinese speakers by handling rare terms effectively.
Authors
Muhammad Huzaifah, Yu Pan, Zachary Yeo, Ningjie Bai, Guangzhao Yang
Abstract
Contextual biasing supplies an ASR system with a list of expected words at inference time, but existing methods rely on word boundaries that Japanese and Chinese do not provide. We present a boundary-free biasing decoder for frozen public CTC models, built on a character-level Aho-Corasick automaton, with no training and no second pass. Two evidence-based mechanisms replace the boundary: a depth-adaptive gate that sets how hard to push from match depth, and reading-space matching for when the audio is right but the characters are wrong. On Aishell-1 NE's hard R1 subset we reach 66.5% recall, above the trained CLAS baseline (64%), transferring to WenetSpeech and to a second architecture without retuning. We release the first open Japanese contextual-biasing benchmark, where biasing lifts rare-word recall by 25 points at precision above 97%, and still by 19 and 22 points against 1,000-word lists.