A Comparative Study on the Biases of Age, Gender, Dialects, and L2 speakers of Automatic Speech Recognition for Korean Language

초록

Recent advancements in the field of Automatic Speech Recognition (ASR) have seen the emergence of large-scale models, contributing to a surge in research and development. The performance of recent ASR models has rapidly improved with the utilization of extensive pre-training datasets. However, challenges persist in enhancing the recognition accuracy for non-mainstream groups such as the elderly and speakers of regional dialects. This paper conducts experiments using Korean speech data to compare and analyze the biases related to gender, age, dialects, and second language (L2) Korean speakers using the Conformer, wav2vec 2.0, and Whisper models. The experimental results showed that female results exhibited better performance in ASR models than those of males, and Whisper exhibited lower biases than two other models in most cases. Furthermore, Whisper demonstrated robustness compared to the other two models in the L2 speakers. Additionally, the analysis of characters with high error rates for each group revealed that, in the case of Korean, spacing and particles exhibited high error rates. It was also observed that characters with high error rates were similar within age groups rather than between gender groups. In this study, we conducted the first-ever examination of various biases in Korean ASR. The identified biases through these experiments may serve as a starting point for research aimed at improving the performance of ASR for non-mainstream groups. This study underscores the significance of addressing biases to advance fairness in the field of ASR.

키워드