Tesis doctoral de Julián David Arias Londoño
Diagnostic and clinical treatment of laryngeal pathologies is currently a problem of great interest for a part of the scientific community related to the digital processing of speech. The main objective in this field of research is the development of computer-aided medical diagnostic tools, enabling an objective assessment of a patient and subsequently improving the diagnosis and clinical treatment given to him/her. Additionally, such systems help to the early detection of diseases that otherwise could remain hidden during a crucial time for a effective treatment. most of the studies that have been conducted in this field are based on linear methods for characterizing the speech signal. Several of the features extracted from such methods have proved to contain useful information for the detection problem. However, several studies have shown that in the speech production process there are different physical phenomena with nonlinear characteristics, which are not characterized by conventional methods based on linear techniques. this thesis is focused on the analysis and characterization of nonlinear components present in speech signals, using state space reconstruction techniques based on the time delay embedding theorem. Their use have been studied as complementary tools to extract information for the automatic detection and grading of pathological voices, and for the automatic assessment of voices according to the grbas quality scale. nonlinear analysis of voice signals is particularly more complex than using other kind of signals, given that during the voice production process there are inherent turbulent events that add random components to the voice signals which, by definition, are not considered by conventional methods of nonlinear analysis. Therefore, this work examines measures conventionally used for the analysis of nonlinear dynamics, as well as various measures of complexity based on information theory which take into account not only the nonlinear components, but also the stochastic components present in time series. Furthermore, there have been proposed three measures of complexity based on stochastic models that allow a better characterization of the state space and provide useful information for the detection system. moreover, this work study different classification schemes for both the problem of voice pathology detection and the multi-class classification problem according to the grbas quality scale. Additionally, it considers the problem of fusinng information from non-linear methods, with noise and cepstral measures, establishing the real capabilities of complexity measures to improve the discrimination of an automatic detection of voice disorders system. Moreover, it provides a methodology fusinng classifiers, yielding a 98.23% ± 0.01 of accuracy in the first case and a 63.56% of correct detection for the automatic grading of voice quality according to grbas scale. the studies performed showed that the error of the automatic detection pathological voices is reduced a 66.67% in comparison to the error obtained using more classic parameterization approaches based on noise measures and cepstral coefficients. In addition, the error of the voice quality gradings according to the grbas scale was reduced a 13.69% comparing to the performance obtained using classic parameterization approaches. These results outperform the best results currently found in the state of the art.
Datos académicos de la tesis doctoral «Stochastic characterization of nonlinear dynamics for the automatic evaluation of voice quality«
- Título de la tesis: Stochastic characterization of nonlinear dynamics for the automatic evaluation of voice quality
- Autor: Julián David Arias Londoño
- Universidad: Politécnica de Madrid
- Fecha de lectura de la tesis: 02/12/2010
Dirección y tribunal
- Director de la tesis
- Juan Ignacio Godino Llorente
- Tribunal
- Presidente del tribunal: pedro Gómez vilda
- mauricio Orozco alzate (vocal)
- montserrat Vallverdú ferrer (vocal)
- katrin Neumann (vocal)