Continuous Speech Recognition of Kannada language using triphone modeling

Abstract

This paper aims to build continuous speech recognition of regional language Kannada using phoneme modeling wherein each phoneme is represented by tristate Hidden Markov Model(HMM) with each state being represented by Gaussian Mixture Model (GMM). The recordings were sampled at the rate of 16 kHz, blocked into overlapped frames of 25msec duration with 10 msec frame overlap using hamming window, converted into 39 dimensional Mel Frequency Cepstral Coefficients(MFCC), then modeled using continuous density HMM. Kannada language has 46 phonemes, out of which 12 phonemes represent vowels (swaragalu) and 34 phonemes represent consonants (vyanjanagalu). The recognition performance is tested for monophone modeling, word internal triphone modeling and tied state triphone modeling for different gaussian mixtures and results have been presented.