Parkinson's disease (PD) is the second most common neurodegenerative disorder after Alzheimer's disease. The main symptoms are tremor, rigidity and movement disorders. Currently there is no cure for PD, but medication and therapy are used to treat its symptoms. As it is idiopathic, meaning in most cases the cause is unknown, it is essential to identify early signs.
Nowadays it is still a new field of study that needs further development. It has been shown that about 90% of PD patients also develop deficiencies in their voice. In the field of research, voice processing has proven to be a powerful tool in the detection of the disease. For these experiments, databases of patients with PD were used.
Recently in the Laboratory of Speech Acoustics a Hungarian speech database has been gathered to analyse and detect PD in speech. The purpose of my thesis was to process these recordings and identify acoustic parameters that may correlate with the severity of PD. For the recordings prosodic features (e.g. pitch, voiced ratio, intensity) and also speech tempo features (e.g. average number of phonemes pronounced per second, average pause ratio) were calculated.
I created different models based on linear regression, support vector machine and neural network for different recording types, then I aggregated their results by speakers as well. The goal was not only to separate healthy participants and PD patients, but also to estimate the severity of PD. The results were promising, to achieve further progress more recordings are needed.