Exploring Gender Disparities in Automatic Speech Recognition Technology

TL;DR

This paper examines how gender representation and pitch variability in training data influence the fairness and accuracy of ASR systems, revealing complex interactions that affect performance across genders.

Contribution

It provides new insights into the impact of training data gender ratios and pitch variability on ASR fairness, emphasizing the importance of data curation.

Findings

Optimal fairness at specific gender distributions

Performance varies with pitch variability

Bias mitigation depends on training data composition

Abstract

This study investigates factors influencing Automatic Speech Recognition (ASR) systems' fairness and performance across genders, beyond the conventional examination of demographics. Using the LibriSpeech dataset and the Whisper small model, we analyze how performance varies across different gender representations in training data. Our findings suggest a complex interplay between the gender ratio in training data and ASR performance. Optimal fairness occurs at specific gender distributions rather than a simple 50-50 split. Furthermore, our findings suggest that factors like pitch variability can significantly affect ASR accuracy. This research contributes to a deeper understanding of biases in ASR systems, highlighting the importance of carefully curated training data in mitigating gender bias.