Assessing Smartphone Speech Recognition Across Diverse English Accents: A Preliminary Study

This study examines the performance of a smartphone-based automatic speech recognition (ASR) system when processing diverse English accents. With the increasing reliance on voice activated artificial intelligence in daily tasks, ensuring equitable ASR performance across linguistic varieties is critical. Using audio data from the CIRCE project corpus, we assess recognition accuracy for eleven English accents selected according to Kachru’s three-circle model (Inner, Outer, and Expanding Circle varieties). Findings highlight disparities in recognition performance and suggest that ASR models exhibit a bias favoring American English (AmE). The study underscores the need for enhanced ASR inclusivity and diversification of training data.