Speaker Independent Isolated Word Recognition based on ANOVA and IFS

Abstract

A crucial step of speaker independent isolated word recognition is to extract meaningful information from speech signal. Speech signal contain meaningful acoustic features and selecting the significant and optimal features set is an important aspect to improve accuracy. This paper proposes a speaker independent isolated word recognition model by selecting optimal number of significant features. The proposed model consists of feature extraction, statistical analysis of feature and feature selection technique. We use 13th dimension of Mel frequency cepstral coefficient (MFCC) to extract acoustic features of speech signal and after that apply statistical analysis, Analysis of Variance (ANOVA) and incremental feature selection (IFS) technique to investigate efficient features and to rank them accordingly. The objective of applying statistical analysis algorithm and feature selection technique on the cepstral feature is to improve the word recognition performance using significant features set. The experimental analysis is carried out using some machine learning techniques such as Artificial Neural Network (ANN), Support vector machine (SVM) and Naive Bayes (NB) classifier. Performance of each individual classifier has been evaluated and described in this paper. From the experimental analysis it has been observed that ANOVA with feature selection technique provide better result for all the classifier as compared to selecting all MFCC feature. The aim of the proposed method is to select minimum significant features set that can improve the recognition rate and reduce the dimension. Recognition accuracy has been achieved using our own recorded English digit database.