Construcción de mapas articulatorios para la detección automática de la enfermedad de Parkinson por medio de la voz

Abstract

Background and objectives

Recent studies have shown that speech analysis provides relevant information to support the diagnosis and monitoring of patients suffering from Parkinson's disease (PD). In this work a methodology is proposed to create articulatory maps based on articulatory and phonological information such that allow a clear and interpretable visualization of the results.

Materials and methods

A total of 100 speakers were recorded while reading a text with 36 words that includes all phonemes of the Colombian Spanish. Phonological features are extracted with two toolkits: PhonVoc and Phonet. Forced alignment is used to obtained the time-stamps per phoneme. Support vector machines and random forests are used to classify between PD patients and non-symptomatic subjects.

Results

Accuracies of up to 90% are observed when the phonological class «Vowels» is considered and also accuracies above 80% are found for «Nasals», «Voiceless ficatives» and «Voiced Stop». Articulatory maps are created based on Gaussian mixture models with the aim to enable the interpretation of results.

Conclusions

The proposed methodology is suitable for the automatic detection of PD and also to assess possible articulatory deficits in the production of specific phonological classes.