generative-bias

VIBE is a framework that evaluates generative bias in Large Audio-Language Models using open-ended tasks with human-recorded speech, revealing systematic biases triggered by gender and accent cues.