LLM-based filtering improves ASR for noisy police audio

Researchers have developed a new method to improve Automatic Speech Recognition (ASR) systems for noisy police audio by using pseudo-labeling. This technique adapts foundation ASR models like Whisper and Qwen3-ASR to specific domains, such as police communications in Baltimore and Chicago. The study found that existing confidence metrics were insufficient for filtering pseudo-labels, leading to the introduction of an LLM-as-a-judge filtering paradigm that significantly reduces Word Error Rate (WER). Additionally, a cross-model pseudo-labeling approach was explored as a promising avenue for future research. AI

IMPACT This research could lead to more accurate transcription of critical communications, improving public safety and operational efficiency.

RANK_REASON The cluster contains an academic paper detailing a new research methodology for improving ASR systems. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Baltimore
• Chicago
• Hugging Face
• Kaavya Chaparala
• Qwen3 ASR
• Whisper

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →