XML Print

Abstract:   (632 Views)
This work presents the development of an emotion recognition system of human speech, using Gaussian Radial Basis function Network (GRBN) trained with the scaled conjugate gradient (SCG) optimization technique. In this study, emotion-based speech feature vectors like MFCC, LPLC and MVDR of both speech and glottal wave signals are extracted, and optimally dimension-reduced using a new pQPSO, then a GRBN classifier which has been trained with the scaled conjugate gradient descent algorithm will classify the emotion of the input speech signal. The SCG belongs to the class of conjugate gradient methods, which illustrates better convergence on most problems and avoids a time-consuming line-search per learning iteration compared to the other conjugate gradient algorithms. On the other hand, Gaussian radial basis network is a widely-used tool for nonlinear function approximation and classification, which is a central theme in pattern recognition. Preliminary results show this new method which has not been used in speech emotion recognition area is more likely to achieve better results in emotion recognition on Berlin Database of Emotional Speech (EMODB) compared to two recent works and other conjugate gradient descent methods.
Full-Text [PDF 1997 kb]   (218 Downloads)    
Type of Study: Research Paper | Subject: Speech Processing
Received: 2017/11/27 | Accepted: 2018/06/06 | Published: 2018/06/06

Creative Commons License
© 2019 by the authors. Licensee IUST, Tehran, Iran. This is an open access journal distributed under the terms and conditions of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.