Bangla Voice Based Assistance for Mobile Device

Abstract

Bangla speech recognition is a relatively young area of research and we have not seen much success so far. Pattern recognition approach is generally used for speech recognition. CMUSphinx is a framework which uses Hidden Markov Model (HMM) for pattern training and n-gram technique to build a language model from the speech corpus which can be handy for building speech recognition systems. The success of speech recognition mostly depends on the speech corpus and a well-trained acoustic model. Such a speech recognizer, implemented in mobile devices, can have tremendous implication on our day to day life. However, to build an efficient acoustic model we need an extensive amount of training data. In this thesis work, we have shown how CMUSphinx can be used to build an acoustic model for Bangla. We have built several acoustic models and tried to improve the accuracy rate. One of our trained models has achieved good accuracy rate. In the latter part of the thesis, we implemented the speech recognizer in Android platform. In this process, we investigated some problems those have to be solved to get comparable accuracy rate in Android. We have also proposed a model for the future continuation of the research.