Classificação de disfonias por meio da análise de medidas não lineares e de quantificação de recorrência

Abstract

In recent years techniques of digital processing of speech signals have been used as an auxiliary tool in the evaluation of vocal deviations, providing the patient with greater comfort low cost and objectivity when compared to the techniques traditionally employed, such as perceptual-auditory analysis. The evaluation of vocal quality, through acoustic analysis of voice signals, is becoming a very popular clinical practice for the detection of vocal disorders that in some cases can be caused by laryngeal lesions or vocal abuse. In this research, we used some traditional non-linear measures combined with measures of recurrence quantification for the discriminative analysis of vocal deviations, breathiness, roughness and strain. The characteristics of the non-linear dynamic analysis,used in the classification process, were the Reconstruction Step (τ), the First Minimum of the Mutual Information Function (PM) and the Correlation Dimension (D2). The quantification measures employed were: Determinism (Det), Shannon entropy (Entr), Mean length of diagonal lines (Lmed), Maximum length of vertical lines (Vmax) and Transitivity (Trans). Through these statistical tests, the potential of each characteristic to discriminate the types of voice signals was evaluated. In the classification process, the neural network MLP (Multilayer Perceptron) was used, with supervised learning algorithm Graded Conjugate Gradient (SCG). There was an average accuracy of 90% in the discrimination between healthy and deviant voices. In the classification between healthy and strained voices, an average accuracy of 76% was obtained with the combined measures Trans, τ , Vmax, Lmed, Det and D2. In the detection of the roughness deviation, an average accuracy of 89% was obtained with the Lmed, Entr, Trans and D2 measures and in the distinction between healthy and breathy voices, 91.17% of accuracy was obtained with only two combined measures, Trans and τ , showing the promising character of the used technique.