Undergraduate thesis · Speech recognition · 2014

Bangla Speech Corpus For Large Vocabulary ASR System

A speech-corpus and baseline recognition study by Anurag Bhattacharjee and Sujoy Datta, built around standard Bangla as spoken across regions of Bangladesh.

Research question

How can a Bangla speech corpus support a large-vocabulary ASR baseline?

The thesis prepared text, recorded audio, phonetic labels, a pronunciation dictionary and phoneme bigrams, then trained and evaluated a recognition system with the CMUSphinx toolchain.

Text + recordingsregional speech samples
Phonetic resourceslabels + dictionary
SphinxTrainacoustic model
PocketSphinxbaseline evaluation
Evaluation

Report the metric with its meaning and scope.

Approximately 39.39%Performance reported in the 2014 thesis. The corpus and evaluation are best understood as an early baseline with limited training data.

What this establishes

Early experience with speech-corpus construction, phonetic resources, acoustic-model training and ASR evaluation.

How it connects to current work

This undergraduate work is a historical foundation for later NLP research; it is not presented as evidence of current production speech-AI experience.