Spoken Stereoset: On Evaluating Social Bias Toward Speaker in Speech Large Language Models
The episode discusses a new paper titled "Spoken Stereoset: On Evaluating Social Bias Toward Speaker in Speech Large Language Models." The authors created a dataset to measure social bias based on voice characteristics like gender and age. They found most models show little bias, though one model overcorrected for age. The discussion concludes that this research opens a path to fix biases specifically within the speech encoding process.