WER and inference-cost benchmark for Indian-English ASR, and an audit of the evaluation itself: 9 systems x 3 corpora, cluster-bootstrap CIs with Holm correction, a reference-artifact taxonomy confirmed by human re-transcription, and a speaker-disjoint fine-tuning study across 6 seeds x 3 sizes.
benchmark reproducible-research evaluation speech-recognition whisper asr model-evaluation nptel wer fine-tuning parakeet huggingface inference-benchmark accented-speech indian-english qwen3 academic-speech aesrc2020
-
Updated
Aug 4, 2026 - Python