Development of Consonant-Vowel Recognition Systems for Indian languages: Bengali and Odia

The basic goal of this work is to develop a Consonant-Vowel Recognition System (CVRS) for determining a sequence of Consonant-Vowel (CV) units present in a given speech utterance. In this work, we are focusing on developing CVRSs for Indian languages namely Bengali and Odia. This framework of developing CVRSs can be extended to any Indian languages. We have developed two separate CVRSs for Bengali and Odia languages. The CVRS is developed using read speech corpus. In this study, 67 CV classes for Bengali and 58 CV classes for Odia are used. Mel Frequency Cepstral Coefficients (MFCCs) are used as features for building models. The Vowel Onset Points (VOPs) are used as anchor points for marking syllable boundaries and for feature extraction. Support Vector Machines (SVMs) are used for building CV models. The performance of the developed CVRSs are evaluated in speaker dependent and speaker independent modes. In speaker dependent case, the best percentage accuracies of Bengali and Odia CVRSs are 49.48 and 69.66 respectively whereas in speaker independent case, the best percentage accuracies of Bengali and Odia CVRSs are 40.26 and 41.59 respectively.