An acoustic model of speech dysprosody in patients with Parkinson's disease

This study aimed to determine the acoustic properties most indicative of dysprosody severity in patients with Parkinson's disease using an automated acoustic assessment procedure.

A total of 108 read speech recordings of 68 speakers with PD (45 male, 23 female, aged 65.0 ± 9.8 years) were made with active levodopa treatment. A total of 40 of the patients were additionally recorded without levodopa treatment to increase the range of dysprosody severity in the sample. Four human clinical experts independently assessed the patients' recordings in terms of dysprosody severity. Separately, a speech processing pipeline extracted the acoustic properties of prosodic relevance from automatically identified portions of speech used as utterance proxies. Five machine learning models were trained on 75% of speech portions and the perceptual evaluations of the speaker's dysprosody severity in a 10-fold cross-validation procedure. They were evaluated regarding their ability to predict the perceptual assessments of recordings excluded during training. The models' performances were assessed by their ability to accurately predict clinical experts' dysprosody severity assessments.

The acoustic predictors of importance spanned several acoustic domains of prosodic relevance, with the variability in f change between intonational turning points and the average first Mel-frequency cepstral coefficient at these points being the two top predictors. While predominant in the literature, variability in utterance-wide f was found to be only the fifth strongest predictor.

Human expert raters' assessments of dysprosody can be approximated by the automated procedure, affording application in clinical settings where an experienced expert is unavailable. Variability in pitch does not adequately describe the level of dysprosody due to Parkinson's disease.