Automatic Detection of Parkinson's Disease from Continuous Speech Recorded in Non-Controlled Noise Conditions

Automatic classification of Parkinson's disease (PD) speakers and healthy controls (HC) is performed considering speech recordings collected in non-controlled noise conditions. The speech tasks include six sentences and a read text. The recording is performed using an open source portable device and a commercial microphone. A speech enhancement (SE) technique is applied to improve the quality of the signals. Voiced and unvoiced frames are segmented from the speech tasks and characterized separately. The discrimination of speakers with PD and HC is performed using a support vector machine with soft margin. The results indicate that it is possible to discriminate between PD and HC speakers using recordings collected in non-controlled noise conditions. The accuracies obtained with the voiced features range from 64% to 86%. For unvoiced features the accuracies range from 78% to 99%. The SE algorithm improves the accuracies of the unvoiced frames in up to 11 percentage points, while the accuracies decrease in the voiced frames when the SE algorithm is applied. This work is a step forward to the development of portable devices to assess the speech of people with PD.