Inclusive ASR for Critical Public Services: Debiasing with Actor-Simulated Speech

Recent advances in automatic speech recognition (ASR) have improved the overall performance of speech recognition, yet regional dialects continue to pose significant challenges. This is particularly critical in public service applications such as legal aid and housing support, where bias in ASR systems inadvertently disadvantages vulnerable groups. While fine-tuning existing models using data from the target application is a common approach to addressing bias, the sensitive nature of these services makes this approach infeasible. To overcome this and ensure inclusivity, we collected over 200 h of actor-generated simulated data, aimed at addressing regional dialects in the United Kingdom, where dialects and accents are interlinked with socioeconomic status. Through a set of rigorous experiments, including fine-tuning several models using simulated data, we demonstrate that simulated data not only improves the real-world performance of models but also provides insights into fine-tuning data configurations that are more effective in practice.