Speech Recognition 🗣️ in Bengali 🇧🇩🇮🇳

Overview

ASR project for Bengali, exploring Automatic Speech Recognition (ASR) techniques.

ASR system transcribes speech into text, emphasizing the probabilistic sequence of words and sounds.

Two approaches: statistical and end-to-end. Acoustic Model and Lexicon play key roles.

I focus here on the end-to-end approach.

Bengali Language

• 09 vowels, 39 consonants
• Official language of Bangladesh
• 2nd most spoken language in India

Spoken by 200M+ people, featuring diverse dialects.

Competition Objective

Handling out-of-distribution data and linguistic nuance, improve WER.

Training Database

• 1180 hours of audio
• Data cleaning: maintain WER < 70%
• Preprocessing: Mel-Log Spectrograms, tokenized embeddings

Models Explored

RRN-T Transducer 🔄

PyTorch implementation with transducer approach.

Whisper by OpenAI 🤫

40% WER on Common Voice, 10% in French.

Results 📚

• Baseline vs. Whisper vs. BenglaASR
• Whisper-small fine-tuned: WER 67% achieve lower WER then the baseline.

More details presentation slides in French here

Sources:

• ASR Overview
• Transducer Implementation
• Speech Recognition in ML