Speech Translation Models Risk Misgendering Speakers
A new study from arXiv (2511.21517) investigates gender bias in speech translation (ST) models. Unlike text, speech carries acoustic cues like pitch that can influence gender assignment when translating from English (a notional gender language) into languages with grammatical gender (Spanish, French, Italian). The research examines how training data, internal language model biases, and acoustic information interact, finding that models do not simply replicate term-specific gender associations. This risks misgendering speakers through masculine defaults or vocal-based assumptions. The study aims to understand the mechanisms behind these decisions.
Key facts
- Study investigates gender bias in speech translation models
- Focuses on English to Spanish, French, and Italian translation
- Acoustic cues like pitch influence gender assignment
- Models risk misgendering speakers
- Training data patterns, ILM biases, and acoustic info interact
- Models do not simply replicate term-specific gender associations
- Published on arXiv with ID 2511.21517
- Study is an interpretability analysis
Entities
Institutions
- arXiv