Developing efficient speech recognition system for Telugu letter recognition

Telugu is the third largest language spoken by nearly 80 million native speakers. Telugu is one of four classical languages in India. Telugu is the official language for the state of Andhra Pradesh. Each telugu word ends with vowels. So there is a scope for research about Telugu vowels recognition rate. The application of machine learning techniques to biometric authentication and recognition problems has gained a widespread acceptance. In this research, both MLP and TLRN models were trained and tested on a dataset that consists of Four different speakers (2Male and 2Female) are allowed to utter the letters for 10 times. Recognition of the telugu letters is carried out in speaker dependent mode. In this mode the tested data presented to the network are same as the trained data. A comparative study of the application of Multilayer Perceptron (MLP) and Time Lagged Recurrent Neural Network (TLRN) in speech recognition has been carried out with the features LPCC and MFCC to obtained spectral and statistical parameters. The goal of speech recognition in biometrics is to verify an individual´s identity based on his or her utterance. It is found that the proposed system outperforms the conventional system with both the features LPCC and MFCC. Promising results are obtained both in the training and testing phases due to exploration of discriminative information with neural networks. It is found that TLRN trains and tests faster than MLP. For both the convention system and proposed system, MFCC gives higher recognition accuracy in training and testing phases. The vowel recognition rate for the convention system with the features LPCC and MFCC are 92.47% and 94% respectively whereas for the proposed system it is 96% and 97.56% respectively.