Temporal Envelope and Fine Structure Cues for Dysarthric Speech Detection Using CNNs

Deep learning-based techniques for automatic dysarthric speech detection have recently attracted interest in the research community. State-of-the-art typically learn neurotypical and discriminative representations by processing time-frequency input such as magnitude spectrum of short-time Fourier transform (STFT). Although these are expected to leverage perceptual cues, STFT do not necessarily convey aspects complex sounds. Inspired temporal mechanisms human auditory system, this paper we factor signals into product a slowly varying envelope rapidly fine structure. Separately exploiting different cues present (i.e., phonetic information, stress, voicing) structure pitch, vowel quality, breathiness), two learned through convolutional neural network used detection. Experimental results show that both yields considerably better performance than only envelope, structure, or representation.