Spectrogram and Mel-Spectrogram Based Dysphonic Voice Detection Using Convolutional Neural Network

A non-invasive pathological voice detection system has been presented in this paper. This work considers two spectral images of voice signals namely spectrogram and Mel-spectrogram to detect dysphonic voices. The spectrogram is a convenient representation of voice signals on a time-frequency scale and has been popularly investigated in pathological voice detection algorithms. However, the spectrogram uses a linear frequency scaling and hence, does not consider the resolution of the human auditory model. The Mel-spectrogram overcomes this limitation by using a quasi-logarithmic frequency spacing. This work uses vowel samples available in the Saarbrucken voice database. The simulation results show that the Melspectrogram outperforms the spectrogram in terms of classification accuracy.