Telugu speech recognition using combined MFCC, MODGDF feature extraction techniques and MLP, TLRN classifiers

Abstract

Telugu is the standard language used to communicate mainly in Andhra Pradesh and Telangana states with approximately 100 million speakers. Every Telugu word ends with vowels. Automatic speech recognition has major applications in the smart world. Telugu speech recognition has a huge scope of research from past decade. This paper deals with Telugu speech recognition in a speaker-dependent format for 10 words which are numbers from one to ten uttered by 10 speakers which creates a data base of 100 samples. Numbers are very frequently spoken words where recognition of these words plays a major role in Telugu speech recognition. These spoken words are preprocessed using various techniques like de-noising, framing, sampling, transformations, and endpoint detection. Methods like mel frequency cepstral coefficients (MFCC) and combined MFCC with modified group delay functions (MODGDF) are used for extracting the features. Extracted features are used for training and testing phase. Multilayer perceptron (MLP) and time lagged recurrent neural network (TLRN) patterns are trained and tested. Comparison is done for MLP and TLRN with the feature extraction techniques MFCC and MODGDF-MFCC. Integrated MODGDF-MFCC gives best accuracy in training and testing phases.