Linear Collaborative Discriminant Regression and Cepstra Features for Hindi Speech Recognition

Abstract

Speech recognition system is one of the significant, but challenging systems in computer-human interaction. Recognizing Indian languages, especially Hindi, find many practical difficulties due to its wide grammatical and phonetic features from English.  This paper focuses on Hindi speech recognition system for which Cepstra features and linear collaborative discriminant regression (LCDR) model are proposed for feature analysis and recognition. For a definite of audio signals, two models of test speech signals are synthesized and experimental investigations are carried out. The performance of the LCDR methods are analysed using Type I and II error functions and compared with the existing methods such as NN₂-cepstra and SVM₂-cepstra. Moreover, the best, worst, mean, median and standard deviation are used for the statistical prediction and the proposed LCDR method is proved as the superior method for recognising Hindi speech.