Two-stage Isolated Consonant-Vowel (CV) Unit Recognition in Indian Languages

This paper addresses the issues in the recognition of the most frequently occurring Consonant-Vowel (CV) units of speech in Indian languages. Two major issues in the recognition of CV units are the large number of CV classes and high similarity among several CV units. In this paper, we propose two level acoustic models comprising of Hidden Markov Models (HMM) and Support Vector Machines (SVM) to recognize 145 CV classes. In the proposed approach CV recognition is done in two stages, first stage vowel will be recognized and in the second stage consonant will be recognized. CV classes are divided into five subclasses based on vowels, to reduce the influence of vowel on CV recognition. Vowel Onset Point (VOP) is used as an anchor point to extract features for building the vowel and consonant recognition models. VOP used in this paper is detected by combining the evidences from excitation source, spectral peaks, and modulation spectrum energies. Proposed VOP based two level HMM and SVM based approach for CV recognition is giving better performance compared to monolithic HMM and SVM.