Comparative Analysis of Multilingual Pre-trained Models for Nepali Automatic Speech Recognition

Author1, Author2, Author3, Author4, Author5

fine-tuningASRmultilingualbenchmarking

2608.12327

Builder Relevance

80%

6h ago

Abstract

This paper benchmarks six multilingual pretrained models for Nepali ASR under a controlled fine-tuning protocol.

Reality Card

Core Claim

Whisper-Large-v3-Turbo and IndicWav2Vec achieve comparable WER despite significant differences in parameters and pretraining data, demonstrating the importance of language-family proximity.

Method / Result

Whisper-Large-v3-Turbo achieves a WER of 14.76%, while CTC decoders run up to 29x faster than autoregressive models.

Limitations

The study's findings are based on a specific Nepali corpus, which may limit generalizability to other languages or dialects.

← Back to all papers