Investigating the Impact of Multilingual Pre-trained Speech Models on Gender Bias in ASR for Low Resource African Languages

Sikasote, Claytone and Suleman, Hussein and Buys, Jan (2025) Investigating the Impact of Multilingual Pre-trained Speech Models on Gender Bias in ASR for Low Resource African Languages, Proceedings of Southern African Conference for Artificial Intelligence Research, 1-5 December 2025, Cape Town, South Africa, Communications in Computer and Information Science, 2784, 144-158, Springer.

Other (PDF) SACAIR_2025_Paper_Camera_Ready-2.pdf - Accepted Version Restricted to Registered users only Download (359kB) | Request a copy

Abstract

While fine-tuning transformer-based pre-trained speech models improves speech recognition for low resource languages, the approach increases the risk of speaker attribute bias in the resulting target language automatic speech recognition (ASR) systems. This work investigates gender bias in two state-of-the-art pre-trained speech models, MMS and Whisper, fine-tuned for ASR on three African languages: Bemba, Nyanja, and Swahili. We fine-tune models on gender-specific as well as gender-balanced datasets, and estimate and compare gender bias across different settings. Our results show varying degrees of gender bias in the fine-tuned models, even with gender-balanced fine-tuning, suggesting influence from pre-trained models. Inconsistencies in gender-specific fine-tuning further confirm the transfer of bias from pre-trained models. Additionally, an ablation study shows no relationship between training data size and gender bias.

Item Type:	Conference paper
Subjects:	Computing methodologies > Artificial intelligence > Natural language processing > Speech recognition
Date Deposited:	02 Nov 2026 07:06
Last Modified:	02 Nov 2026 07:06
URI:	https://pubs.cs.uct.ac.za/id/eprint/1763

Actions (login required)

View Item