Evaluating OpenAI’s Whisper ASR: Performance Analysis Across Diverse Accents and Speaker Traits

This research explores the performance of the Whisper's ASR system on different native and non-native English accents. The findings indicate better performance on North American vs British and Irish English accents; and on native vs native accents. The analysis also unearths links between speaker traits (sex, L1 typology, and L2 proficiency) and word error rate. An unsupervised K-means analysis identified ten distinct clusters in the data, providing valuable insights into the relationship between speaker characteristics and ASR performance. Additionally, the study found that Whisper performed better on read speech than on conversational speech. The implications of these findings are discussed.