Choo , Chian Choong
(2016)
Hybrid Mfcc And Lpc For Stuttering Assessment Using Neural Network.
Masters thesis, Universiti Sains Malaysia.
Abstract
Stuttering is characterized by disfluencies, which disrupt the flow of speech. Traditional way of stuttering assessment is time consuming. The stuttering assessment results always inconsistent between different judges, because human perception on the stuttering event are different for each individual. The stuttering assessment system will reduce the tedious manual work and improve the consistency of the assessment result. The objective of this project is to develop classifier for prolongation and repetition disfluencies in speech using artificial neural network. Three different feature extraction was used in this project, which is Mel Frequency Cepstral Coefficient (MFCC), Linear Prediction Coefficient (LPC) and hybrid MFCC and LPC. The flow of the project were: 1) Stuttered speech data acquisition; 2) Word segmentation and categorization; 3) Feature extraction using 3 different methods; 4) Classification using neural pattern recognition in Matlab. The overall accuracy of the 3 different feature extraction used were 84.6% (LPC), 84.6% (MFCC) and 88.5% (hybrid MFCC and LPC). The classification accuracy using hybrid MFCC and LPC with respect to target classes, which were prolongation, repetition and fluent, were 66.7%, 92.3% and 96.3%. A disfluencies classifier had been developed with hybrid MFCC and LPC as feature extraction and ANN as a classifier. The overall performance of the disfluencies classifier is 88.5%.
Actions (login required)
|
View Item |