RedVox: Safety and Fairness Gaps in Speech Models Across Languages

RedVox introduces a multilingual safety and fairness benchmark for audio and speech models, built from real voices and covering English, French, Italian, Spanish, and German. The authors report that only 8% of surveyed state-of-the-art speech model releases document multilingual analysis. Evaluations of eight models indicate that vulnerabilities remain under non-adversarial conditions, become worse in non-English languages, and are amplified when requests are spoken rather than provided as text. The paper also examines privacy and personal challenges in collecting naturalistic speech data.