LLM filter trims errors in AI transcriptions of noisy police audio

The team led by Kaavya Chaparala evaluated two foundation speech‑recognition models—OpenAI’s Whisper and Qwen3‑ASR—on noisy broadcast police communications from Baltimore and Chicago. They found the models’ built‑in confidence scores could not separate good from bad machine‑generated transcripts. By using a large language model as an external judge to discard context‑implausible pseudo‑labels, they cut the word‑error rate of the training data. The paper, accepted to the SLT 2026 conference, also proposes swapping pseudo‑labels between models as a future direction.