Adaptive context biasing in transformer-based ASR systems

With the advancement of neural networks, end-to-end neural automatic speech recognition (ASR) systems have demonstrated significant improvements in identifying contextually biased words. However, the incorporation of bias layers introduces additional computational complexity, requires increased resources, and leads to redundant biases. In this paper, we propose a Context Bias Adaptive Model, which dynamically assesses the presence of biased words in the input and applies context biasing accordingly. Consequently, the bias layer is activated only for input audio containing biased words, rather than indiscriminately introducing contextual bias information for every input. Our findings indicate that the Context Bias Adaptive Model effectively mitigates the adverse effects of contextual bias while substantially reducing computational costs.