A Novel Approach for Multilingual Speech Recognition with Back Propagation Artificial Neural Network

Abstract

“Speech Recognition” of audio signal is important for telecommunication, language identification and speaker verification. Robust Speech Recognition can be applied to automation of houses, offices and telecommunication services. In this paper Speech Recognition & Language Identification have done for Bengali, Chhattisgarhi, English and Hindi speech signals. The Bengali, Chhattisgarhi, English, Hindi speech signals are “Ekhone Tumi Jao”, “Ae Bar Teha Ja”, “Now This Time You Go” and “Ab Is Bar tum Jao” respectively. This method is mainly applied in two phases, in the first phase Speech Recognition and Language identification have done with Back Propagation Artificial neural Network (BPANN) and in the second phase Speech Recognition and Language Identification have done with the combination of the Particle Swarm Optimization (PSO) feature selection technique and BPANN. For the feature extraction Mel Frequency Cepstral Coefficients (MFCC) & Linear Predictive Coding (LPC) is used. MFCC and LPC are the most widely used feature extraction method. BPANN is a feed forward type neural network, it can trace back the error signal for weight modification, error signal generates when the actual output value differs from the target output value. The system accuracy and performance is measured on the basis of “Recognition Rate” and amount of error. Multilingual Speech Recognition and Language Identification with PSO feature selection technique gives the better Recognition Rate as compare to the without PSO feature selection technique.