researchMd. NayeemOct 8, 2026
Low-Resource Acoustic Modeling for Regional Bengali Dialects
A comparative evaluation of self-supervised speech representations (Wav2Vec 2.0 & HuBERT) for low-resource Bengali speech recognition.
Abstract
Regional dialectal variations in the southern regions of Bangladesh present acute acoustic recognition challenges. We construct a multi-dialect acoustic dataset and evaluate fine-tuning dynamics on multilingual Wav2Vec 2.0 representations.
Experiments & Findings
Cross-dialect fine-tuning reduces Word Error Rate (WER) by 14.2% across conversational benchmarks.