Towards Inclusive and Fair ASR: Insights from the SAPC Challenge for Optimizing Disordered Speech Recognition

ASR has advanced significantly, yet remains limited for impaired speakers due to data scarcity. In response to this gap, the Speech Accessibility Project (SAP) represents a significant initiative in data collection on impaired speech. This paper reports our participation in the SAPC challenge, where we leveraged SAP data to improve ASR performance for disordered speech. Our system ranked fourth in terms of Word Error Rate, recording values of 10.06% and 11.8% WER for Test1 and Test2 subsets, respectively, on the challenge leaderboard. In particular, our research examines the power of SOTA ASR models to capture contextual information in the presence of disordered speech disfluencies. We focused on two ASR architectures, ContextNet and Parakeet, based on their documented ability to efficiently and effectively handle contextual information for typical speech, utilizing distinct mechanisms. Our experiments demonstrated that Parakeet slightly outperformed ContextNet, as evidenced by WER.