Comparing Different Pattern Recognition Approaches of Building Marathi ASR System

Abstract

This paper investigates the efficiency of various pattern recognition approaches for building Marathi ASR system. It utilizes Mel Frequency Cepstral Coefficients (MFCC) feature extraction technique for extracting features from acoustic signals. In this paper, pattern recognition experiments are carried out on dataset having different speakers, environment and location. This paper compares different approaches such as KNN, SVM, NN, DNN and DBN which can be utilized for building speech recognition system. DBN obtained comparatively good results while compared using different algorithm analysis parameters. The algorithms are analyzed using two measures i.e. positive measures which is having 7 parameters whereas negative measure is having 3 parameters. DBN obtained maximum accuracy score of 81.48% in pattern recognition.