AI
Pretrained ASR Pseudo-labeling for Noisy Police Audio
Publication: cs.AI updates on arXiv.org
Author: Kaavya Chaparala, Su Huang, Stephen L. Miller, Rhiannon N. Miller, Anjalie Field
Source published: 28/09/2026, 04:00:00 UTC
Briefing created: 28/09/2026, 04:11:05 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 28/09/2026, 04:03:24 UTC.
Briefings
This study evaluates pseudo-labeling for improving ASR on Baltimore and Chicago police communication data. Researchers found that existing internal confidence metrics cannot reliably distinguish high-quality from low-quality transcripts, leading to significant word error rates. They introduced an external LLM-as-a-judge filtering paradigm that successfully reduces these errors compared to internal metrics, though a performance gap remains. The study also proposes a cross-model pseudo-labeling approach where one model is fine-tuned using labels from another, identified as a promising future direction.
Key points
- Existing internal confidence metrics (log-probabilities and STAR scores) fail to distinguish between high and low quality BPC pseudo-labels.
- An external LLM-as-a-judge filtering paradigm was introduced to discard contextually implausible transcripts, significantly reducing WER across Baltimore and Chicago BPC corpora.
Context and limitations
The study covers only Baltimore and Chicago BPC corpora; no other domains were tested. Results are based on preprint arXiv:2609.30469v1 and remain subject to future validation. Performance gaps relative to an oracle filter are substantial but unmeasured.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 28/09/2026, 04:11:05 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeifautov3rjv6n5i7xrz6unp5n5h4ucbxayxe2l3bk3wdlhf377n6e