Malayalam Speech Recognition system and its application for visually impaired people

Abstract

This paper describes the development of state-of-the-art large vocabulary continuous speech recognition (LVCSR) system for the Malayalam language with an application for visually challenged. For an LVCSR, building a high accurate acoustic models and large-scale language models are the challenging task. Speech corpus for training the system is collected from 80 native speakers in room environment ensuring the speaker variance. Mel-frequency Cepstral Coefficients (MFCC) method is used as a front-end to extract acoustic features from the input signal. Acoustic model is built on 30 hours of speech data based on Hidden Markov Model (HMM). A hybrid model, integrating rule based and statistical method is used to handle pronunciation variations in the dictionary. The best configuration of the system achieved word accuracy of 75% in average. Accuracy of the system is further increased up to 80% in average, by implementing speaker adaptation technique. The developed system is integrated to OpenOffice Writer together with TTS for making it user friendly editor for visually challenged people.