Turkish vowel classification based on acoustical and decompositional features optimized by Genetic Algorithm

Abstract

In digital speech processing (DSP), vowel classification is one of the important and challenging issue of automatic speech/speaker recognition (ASR) systems where vowels need to be processed in order to identify spoken speech or speaker identity. Because the vowels' acoustic features such as Formant Frequencies, Energy, Zero Crossing Rate (ZCR) etc. contain patterns revealing speech/speaker characteristics, vowels were used in many studies like speech/speaker recognition, speaker verification, accent/dialect estimation, illness detection via speech etc. In this work, digital speech signals of 8 vowels (/a/, /e/, /1/, /i/, /o/, /o/, /u/ and /u/) in Turkish language were classified using a feature set optimized by Genetic Algorithm. The entire feature vector, which has length of 24, comprises acoustical features (Formant Frequency obtained by Linear Predictive Coding-LPC, Energy, ZCR, Mel-Frequency Cepstral Coefficients-MFCCs) and decompositional features (Wavelet Decomposition Shannon Entropy). All of the 8 isolated vowels were uttered by 10 male speakers (only male to avoid gender effect), and totally 2762 observations were obtained from the vowels signals for classification step. The results inferred that the feature vector optimized by Genetic Algorithm method reached 100% classification success accuracy via the simple 1NN classifier with Manhattan distance. These findings proved success of the proposed model and the Genetic Algorithm's efficiency on feature selection for the Turkish vowel classification. (C) 2019 Elsevier Ltd. All rights reserved.