Description of the experimental set up.

The selected benchmark features correspond to the ones of [29], i.e. MFCCs, MGDCCs, Jitter(%): frequency perturbation, Shimmer (dB): amplitude perturbation, APQ (%): amplitude perturbation quotient, PPQ (%): pitch perturbation quotient, RAP (%): relative average perturbation, CPP mean: mean of cepstral peak prominence corresponding to the mean of voice quality perturbation and CPP s.d.: variation in the cepstral peak prominence corresponding to variation in voice quality perturbation. These are extracted on the given speech signals . The configuration employed for the extraction procedure of these features are provided in subsection 7.4. Then, each system model is performed, and the GLRT is applied. Note that, SM1 is considered as benchmark model since it is the proposed reference given standard ASR direclty extract features on the raw data (as done for the bencmark introduced). Further, when it comes to SM2 and SM3, we will consider the first three IMFs or the first three BLIMFs only since they are the ones that detect the great majority of formants required for the classification of Parkinson’s disease. Both the SVM and the GLRT will be done by patient, setting up a text-dependent and a speaker-dependent environment.