The Voice Behind the Words: Quantifying Intersectional Bias in SpeechLLMs
A new research paper has quantified intersectional bias in Speech Large Language Models (SpeechLLMs). The study used 2,880 controlled interactions across six English accents and two gender presentations, employing voice cloning to maintain consistent linguistic content. Results indicate that Eastern European-accented speech, particularly from female-presenting voices, receives lower helpfulness scores, though politeness remains consistent. While LLM judges detected these biases, human evaluators demonstrated higher sensitivity to accent-based differences. AI
IMPACT Highlights the need for more nuanced bias detection in speech-based AI systems, particularly concerning intersectional factors like accent and gender.