Predictability Measurements Using Wavelet-Packets and Fractal Modeling for Larynx Pathology Differentiation

• Grant number: 09/10457-2
• Support Opportunities: Scholarships in Brazil - Post-Doctoral
• Start date: December 01, 2009
• End date: May 31, 2010
• Field of knowledge: Engineering - Electrical Engineering
• Principal Investigator: Rodrigo Capobianco Guido
• Grantee: Paulo Rogério Scalassara
• Host Institution: Instituto de Física de São Carlos (IFSC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• Associated research grant: 05/00015-1 - SpeechAuth: development of a biometric authentication and comand interpretation system based on speech processing, AP.JP

Abstract

Considering the point of view of systems, one can state that voice signals carry all intrinsic information of a patient's vocal apparatus, therefore, several characteristics of his or her vocal tract can be obtained from these signals. During previous works, we showed results of voice analysis using predictability measures such as entropy, relative entropy, and predictive power, aiming at differentiating healthy, nodule-affected, and Reinke's edema signals. In this post-doctoral study, we intend to use a new model based on wavelet-packet decomposition and Shapelet transform. This transform uses measures based on fractal analysis, which considers auto-similarities among voice signals. This model will be adopted in order to improve autoregressive and traditional Daubechies' wavelet decomposition models previously used. We also intend to apply a technique called predictable component analysis, which has only recently been started and has the ability to decompose voice signals in components of different degrees of predictability, allowing signal reconstruction with only the most predictable components. Previously, we performed some tests using voice signals obtained from EESC/USP Signal Processing Laboratory database, which was built alongside with the otorhynolaringology sector of the Hospital of Clinics of the Medicine School at Ribeirão Preto (FMRP/USP). Such signals presented problems due to many acquisition and diagnosis differences, requiring several preprocessing and co-analysis steps. In view of this, only a few of them could be applied to our study. In this project, the intention is to use a new voice database provided by the Hospital of Clinics of the University of Iowa (UIHC), which contains dozens of voice signals of many pathologies. With such database, we will expand the results obtained during the doctorate for only two pathologies, with the aid of a more modern and efficient model. This research is inserted in the project SpeechAuth supported by FAPESP, process number 05/00015-1.