Beyond WER: Entity and Disfluency Recall in Accented Conversational ASR
A three-stage pipeline for speakers from India, Indonesia, and Latin America achieves 80-85% entity recall, 76-86% filler recall, 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.
Fiza Husain, Ankit Pandey, Y. Singh
· 0 citations