IndicWav2Vec Matches Whisper-Large-v3-Turbo Performance for Nepali ASR
August 14, 2026
Fine-tuning on 165 hours of Nepali data shows IndicWav2Vec (14.89% WER) performs nearly identically to Whisper-Large-v3-Turbo (14.76% WER). The results suggest that language-family pretraining proximity can compensate for a 9x parameter gap and 40x pretraining-data gap.
HOW THIS AFFECTS YOU
●
builderYou can deploy much smaller, more efficient models for Nepali ASR without significant loss in accuracy.
●
researcherThis provides empirical evidence for the effectiveness of language-family-based pretraining over raw parameter scale.