Gender Bias in Spoken Language Translation: A Review

Abstract

Deep neural network architectures for speech have been found to learn representations that contain information about the speaker, such as the gender. Current speech-to-speech translation technologies often employ a series of steps in cascade, comprising automatic speech recognition, machine translation, and text-to-speech synthesis. Recent studies have discovered that many speech translation systems have a gender bias. This review focuses on research conducted between 2020 and 2025 on gender bias, as speech technology has evolved significantly over the past five years. Our review considers how this bias arises, spreads and continues in systems of speech translation. The sources of bias can be found at various levels of the system, encompassing the stages of data collection, representation, and how the model learns from data, as well as evaluating the performance of the ST system. Gender bias in the datasets is reviewed. Techniques to reduce bias are outlined. While much of the existing research into gender bias has been conducted in European languages, a lot of attention is being paid to languages in which there are fewer resources

languages spoken in India are a focal point of research into gender bias studies. Our review concludes by outlining open research directions and advocating for gender-aware, culturally grounded, and ethically informed approaches to the design and evaluation of future speech translation systems. © 2026 IEEE.