speech AI
Why One Training Run Hides Bias in Speech LLMs
New research reveals that fairness evaluations of adapted speech LLMs can swing dramatically based on a single random training seed — exposing a reproducibility blind spot that affects how we trust voice and audio AI systems.