BRANDT, R. R.; http://lattes.cnpq.br/9572151748158678; BRANDT, Robert Rice.
Resumo:
Voice is very important in human communication, but pathologies can hinder this
communication. The identification of pathologies is an important step for any treatment.
Acoustic analysis is a non-invasive technique that can help in pathologic diagnosis. Voice
signal acoustic analysis is used in this thesis to classify voices as healthy or pathological,
using parametric and non-parametric features of the voice signal. The non-parametric features
include frequency perturbations (jitter), amplitude perturbations (shimmer), and their
variations, along with the presence of glottal noise. The parametric features include Linear
Predictive Coding (LPC) and cepstral analysis. This thesis aims to classify voices as normal
or pathological, and in the pathological group checks for Edema and Paralysis. Results
indicate that, using parametric and non-parametric features, classification between normal and
pathological can be achieved with an accuracy rate of 94.6% for male voices and 87.3% for
female voices. In checking for edema, the accuracy rate was 77.1% for male voices and
54.3% for female voices. In checking for paralysis the accuracy rate was 60.0% for male
voices and 68.6% for female voices. In discriminating between voices affected by edema and
by paralysis, the accuracy rate obtained was 73.5% for male voices and 62.5% for female
voices. The use of a hybrid vector composed of parametric and non-parametric features to
classify voices as normal or pathological, is very promising. On the other hand, the
classification between pathologies did not improve the parametric feature classification.