Análise da relevância das faixas de frequência na detecção de distúrbios vocais por meio de redes neurais convolucionais em espectrogramas.

Abstract

The use of non-invasive techniques for assessing vocal quality has become increasingly common due to recent advances in signal processing and machine learning (ML), particularly in the context of detecting vocal disorders. Among these techniques, spectrogram analysis is widely used to assess spectral patterns, typically limiting analysis to regions below 5.000 Hz, whether through visual inspection or ML algorithms. In this study, we sought to evaluate the relevance of different frequency bands in classifying healthy and disordered voices using convolutional neural networks (CNNs). Additionally, we investigated whether combining results across different frequency bands could enhance classification performance. To achieve this objective, data subsets containing spectrograms from 16 frequency bands were generated through a bank of band-pass filters. This process was applied to two voice datasets. The spectrogram data subsets were used to train 16 pre-trained CNN models, one for each frequency band. The study initially evaluated the relevance of each frequency band individually, followed by an analysis of the results across the 65,536 possible combinations of the 16 frequency bands. This analysis revealed that it is possible to identify patterns of vocal pathologies in frequency regions above 5.000 Hz within the assessed range from 1 to 12,427 Hz. However, the interval between 1 and 1,462 Hz demonstrated substantially higher descriptive capacity. Furthermore, it was observed that combining results from models trained on different frequency bands yields improved classification outcomes, increasing the descriptive capacity of the model, as measured by the Matthews Correlation Coefficient, from 0.6182 to 0.6514 on the SVD dataset, and from 0.5683 to 0.6423 on the AVFAD dataset.