Ranking and Selection of Bias Words for Contextual Bias Speech Recognition

Contextual Automatic Speech Recognition (ASR) systems have made significant advancements. However, contextual ASR models face challenges when dealing with a large number of bias words. This paper focuses on addressing the limitations of contextual ASR models in handling a substantial number of bias words. First, to guide the model to focus on the most important words, we propose a novel network serving as a scorer for bias word ranking and selection. Second, as an example, we explore the use of the proposed scorer in conjunction with the contextual Whisper model. We create a new bias word list using a named-entity recognition (NER) model, which is closer to real-world scenarios. The results on the LibriSpeech dataset with the IS21 bias words list demonstrate that bias word ranking and selection can significantly enhance the model's performance in recognizing bias words, achieving a relative reduction of over 40% in the Biased Word Error Rate.