Mitigating Gender Bias in Accent Recognition Through Gender-Specific Models: An Analysis of Performance and Fairness

Accent recognition plays a crucial role in machine learning and natural language processing. However, since most algorithms in the field of fair machine learning cannot directly work on audio data, therefore research on fairness in accent recognition remains insufficient. This paper investigates gender bias in accent recognition. 760 English audio files spoken by native English, Arabic, and Spanish speakers were utilized, with an equal distribution of 380 male and 380 female speakers. Each audio file was transformed into a spectrogram, and a bag-of-visual-words method was used to convert the spectrogram into vector of feature frequencies. These vectors served as inputs for random forest and K-nearest neighbor. The female-only model exhibited an average improvement of 5.34^%, with a peak performance increase of 8.91% compared to the gender-agnostic model. In a similar vein, the male-only model demonstrated an average performance boost of 4.97%, with a maximum improvement of 8.25%. Crucially, the adoption of gender-specific models resulted in a reduction of gender bias from an average of 10.35% down to 3.83%, substantially diminishing the performance discrepancies between genders.