The Voice Behind the Words: Quantifying Intersectional Bias in SpeechLLMs

The study presents a large-scale evaluation of intersectional bias in Speech Large Language Models (SpeechLLMs), analyzing 2,880 interactions across six English accents and two gender presentations. It finds that Eastern European-accented speech, especially from female-presenting voices, receives lower helpfulness ratings, indicating a significant bias in model performance based on speaker identity. This research highlights the importance of addressing these biases in SpeechLLMs to improve fairness and effectiveness in AI applications, particularly in user-facing technologies.