Representation learning strategies to model pathological speech: Effect of multiple spectral resolutions

Gabriel Figueiredo Miller, Juan Camilo Vásquez-Correa, Juan Rafael Orozco-Arroyave, Elmar Nöth

Abstract: Highlights•A novel auto-encoder fusion strategy is proposed to assess Parkinson’s dysarthria.•Strategy leverages varied spectral resolutions of speech as basis for a feature set.•Framework fuses bottleneck features and reconstruction error of varied spectra.•Proposed approach classifies Parkinson’s with 97% accuracy using cross-validation.•Output highly correlates with clinically assessed dysarthria ratings (ρ=0.79).