SHOBDOTORI: Regional-to-Standard Bangla Speech Recognition

Developed an Automatic Speech Recognition (ASR) platform that transcribes diverse regional Bangladeshi dialects into standard Bangla. Leveraging transformer-based speech models, phoneme alignment, audio augmentation, and n-gram post-processing, it achieves high transcription accuracy and competitive performance on benchmark evaluations.

View Code

SHOBDOTORI: Regional-to-Standard Bangla Speech Recognition — screenshot 1 of 4

SHOBDOTORI: Regional-to-Standard Bangla Speech Recognition — screenshot 2 of 4

SHOBDOTORI: Regional-to-Standard Bangla Speech Recognition — screenshot 3 of 4

SHOBDOTORI: Regional-to-Standard Bangla Speech Recognition — screenshot 4 of 4

1 / 4

Key Features

Automatic transcription of 20 regional Bangladeshi dialects into standard Bangla
Curated dataset with 3,800+ audio samples
Transformer-based models: Whisper and Wav2Vec2
Phoneme alignment and audio data augmentation for robust learning
N-gram KenLM post-processing for fluent and linguistically correct transcriptions
Performance evaluation using Normalized Levenshtein Similarity (NLS)
AI Hackathon project under Televerse 1.0, CUET (Department of ETE)

Challenges Faced

Regional Bangla dialects have limited digitized speech data, making training data scarce
Phoneme-level differences between dialects and standard Bangla caused frequent transcription errors
Balancing transcription fluency with linguistic accuracy

Solutions Implemented

Curated and augmented a dataset of 3,800+ audio samples across 20 regional dialects
Fine-tuned Whisper and Wav2Vec2 with phoneme alignment for dialect-aware transcription
Added an n-gram KenLM post-processing pass to smooth output into fluent, correct standard Bangla