A new paper investigates how large language model (LLM) decoders impact fairness in speech recognition systems. Researchers found that LLM decoders do not necessarily amplify racial bias, with one model even showing improved ethnicity fairness. However, certain models like Whisper exhibited significant issues with specific accents and under acoustic degradation, sometimes leading to pathological hallucination or repetition loops. The study suggests that audio encoder design, rather than LLM scale, is more critical for achieving equitable and robust speech recognition. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
RANK_REASON The submission is an academic paper on arXiv evaluating LLM decoders for bias in speech recognition.
