Beyond WER: Entity and Disfluency Recall in Accented Conversational ASR

A three-stage ASR pipeline built on Qwen2.5-Omni-3B achieved 80-85% entity recall and 76-86% filler recall across 6,000 test utterances from speakers in India, Indonesia, and Latin America, up from 53-55% and under 5% respectively, according to the arXiv paper 2609.20828v1. The pipeline combines heuristic SQL filters that curate entity-rich training data at 2.8x the entity density of random sampling, regional LoRA adapters that emit verbatim and corrected transcripts in a single forward pass, and a six-category error taxonomy validated by an LLM-based judge at 83.8% agreement on 210 human-labelled samples. It reached 6-10% WER, outperformed Whisper and a commercial ASR on entity recall, and matched a zero-shot 30B model with 10x fewer parameters, with paired bootstrap tests attributing 2.8-4.2 percentage points of entity recall gain to curation alone (p<0.0001).

arXiv:2609.20828v1 Announce Type: new Abstract: ASR systems optimised for Word Error Rate WER often miss named entities and filled pauses in accented conversational English, both critical for language-learning feedback. We present a three-stage pipeline for speakers from India, Indonesia, and Latin America: 1 heuristic SQL filters curating entity-rich training data at 2.8x the entity density of random sampling, 2 regional LoRA adapters fine-tuned on Qwen2.5-Omni-3B producing both verbatim and corrected transcripts in a single forward pass, and 3 a six-category error taxonomy validated by an LLM-based judge 83.8% agreement, 210 human-labelled samples . The pipeline achieves 80-85% entity recall up from 53-55% , 76-86% filler recall up from <5% , and 6-10% WER across 6k test utterances, outperforming Whisper and a commercial ASR on entity recall while matching a zero-shot 30B model with 10x fewer parameters. Paired bootstrap tests confirm that curation alone accounts for 2.8-4.2 pp of entity recall gain p<0.0001 .