Performance Analysis of Isolated Speech Recognition SystemUsing Kannada Speech Database

In this article, performance analysis of speech recognition system for different acoustical models has been presented. In the present work, one of the well-known south Indian language named "Kannada" language is considered. Significantly large amount of work has been reported for Automatic Speech Recognition (ASR) in European languages whereas quite a small number of publications can be found in Indian languages. One of the reasons for this gap is that standard speech database in Indian languages is not available. In this study, Kannada speech corpus based on Kannada broadcast news data has been developed. The isolated speaker independent speech recognition system has been developed using Hidden Markov Tool Kit (HTK). The system front-end uses Mel frequency cepstral coefficients (MFCC) and its derivatives as acoustic features whereas acoustical models are developed by using Hidden Markov Models (HMM). Syllable and mono-phone based Kannada dictionaries have been developed in this study. Various mono-phone models considered in this work are word-level, syllable-level and phone-level models. Further, performance evaluation of mono-phone and tri-phone acoustical models for large sized dictionary also carried out. The best word recognition accuracies of 67.82% and 70.56% are reported for mono-phone and tri-phone based systems respectively. The recognition results for different HMM based acoustical models are obtained and hence the recognition performance has been analyzed.