Gated Adapters with Balanced Activation for Effective Contextual Speech Recognition

In end-to-end (E2E) automatic speech recognition (ASR), ac-curately recognizing rare words, such as named entities, re-mains a significant challenge. Although existing contextual biasing techniques have improved recognition rates for named entities, they often incur substantial computational costs and the risk of false biasing. Recent research has shown that integrating gating mechanisms with contextual biasing adapters can dynamically regulate activation, effectively re-ducing unnecessary computational overhead. However, we observed that gating mechanisms tend not to activate when encountering particularly rare instances within named entities. To address this challenge, we combined the gating mecha-nism with a novel activation-balanced objective, resulting in the gate-balanced adapter. This approach not only sustains high recognition rates for named entities but also significantly reduces character error rates (CER) and overall computational load. A series of experiments were conducted on the AISHELL-1 dataset, and the results showed approximately a 1.2% reduction in CER compared to the baseline, highlighting its potential for practical applications.