Phonetic analysis and deep neural network examination for dysarthria detection beyond sociolinguistic differences: a comparison

Abstract

The goal of the paper is to compare a phonetic and an automatic evaluation, via Convolutional Neural Networks, of both dysarthric and control speech. On the one hand, we want to observe if and how phonetic and automatic classifications of dysarthric and healthy speech perform differently, given their specificities; on the other hand, we aim at checking if the automatic analysis is deceived by sociophonetic traits, taking them as instances of non-precise speech production. For the purpose of the analysis, we used a corpus of acoustic recordings of speech produced by 15 male parkinsonian dysarthric speakers (8 from Bari; 7 from Lecce) and 10 male control speakers from the same areas. Part of the corpus was used for fine-tuning the Neural Network that was previously trained on Spanish data; the remaining part of the corpus was used for testing the automatic and the phonetic evaluation. The comparison shows that the automatic and phonetic analyses seem to offer similar results in terms of classification, also with respect to sociolinguistic features. However, they could be successfully integrated as they suggest different information on the speakers and their speech.