Feature extraction and classification of malay speech vowels

In human language, a phoneme is the smallest structural unit that distinguishes meaning. Normally, language like English commonly combines phonemes to form a word. In many languages, the Consonant-Vowel (CV) units have the highest frequency of occurrence among different forms of subword units. Th...

Full description

Saved in:
Bibliographic Details
Main Author: Shahrul Azmi, Mohd Yusof
Format: Thesis
Language:English
Subjects:
Online Access:http://dspace.unimap.edu.my:80/xmlui/bitstream/123456789/12916/1/p.%201-24.pdf
http://dspace.unimap.edu.my:80/xmlui/bitstream/123456789/12916/2/Full%20Text.pdf
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In human language, a phoneme is the smallest structural unit that distinguishes meaning. Normally, language like English commonly combines phonemes to form a word. In many languages, the Consonant-Vowel (CV) units have the highest frequency of occurrence among different forms of subword units. Therefore, recognition of CV units with a good accuracy is crucial for development of a speech recognition system. There are also many applications that use vowels phonemes. Among them are speech therapy systems that improve utterances of word pronunciation especially to children. There are also systems that teach hearing impaired person to speak properly by pronouncing words with a good degree of intelligibility. All of these systems require high degree of vowel recognition capability in which this study focuses on. This thesis contributes five modified feature extraction methods for vowel recognition based on intensities of the Frequency Filter Bands. They are First Formant Bandwidth (F1BW), Fixed Formant Frequency Band (FFB), Spectral Delta (SpD), Bark Intensity (BrKI) and Formant Frequency Difference (FFD). The performance of these five proposed methods are compared with performance of three conventional feature extraction methods of single frame Mel-frequency cepstrum coefficients (MFCCs), multiple frame Mel-frequency cepstrum coefficients (MFCCf) and the first three formant features. The classifiers analysed in this study were Multinomial Logistic Regression (MLR), Levenberg-Marquardt (LM) network, k-Nearest Neighbors (KNN) and Linear Discriminant Analysis (LDA). There are four main contributions of this thesis. First is the new vowel corpus consisting of more than 1300 recorded vowels from 100 Malaysian speakers. Second are the five improved feature extraction methods which perform better than MFCC on single frame analysis. The third is the performance and robustness analysis using different classifiers and different Gaussian noise level. The fourth contribution is the frame analysis criteria for isolated vowel analysis.