a. The duration is the total duration (all the occurrences cumulated for one type of task) and is an average among the participants b. DDK stands for… [601451]

Institut Mines-Télécom, Fondation Télécom, Institut Carnot Telecom & Société Numérique Automatic Detection of Early stages of Parkinson's Disease throw Acoustic Voice Analysis with Mel-Frequency Cepstral Coefficients L. Jeancolas H. Benali B.-E. Benkelfat M. Vidailhet S. Lehéricy D. Petrovska-Delacrétaz affiliations: to be completed Abstract— Vocal impairments are one of the earliest disrupted modalities in Parkinson's disease (PD). Most of the studies whose aim was to detect Parkinson's disease throw acoustic analysis use long-term acoustic parameters. In the meantime, in speaker and speech recognition, analyses are carried out by short-term parameters, and more precisely by mel-frequency cepstral coefficients (MFCC). This paper presents a transposition of the methodology used in speaker recognition to the detection of early stages of Parkinson's disease. Automatic analyses were performed during 4 tasks: sustained vowel, fast syllable repetitions, free speech and reading. Men and women were considered separately in order to improve the classification performance. A leave one out cross validation exhibits accuracies ranging from 60% to 91% depending on the speech task and on the gender. Best performances are reached during the reading task (91% for men). This accuracy is in line with the best classification results in early PD detection found in literature. Keywords— Parkinson's Disease; automatic detection; mfcc; speech signal processing; classification I.   INTRODUCTION Parkinson's disease (PD) is a neurodegenerative disease whose prevalence increases with age. It affects 1% of the people older than 60 years old and up to 4% of those over 80[1]. It is the second most common neurodegenerative disease after Alzheimer's disease. Symptoms are mainly motor symptoms and results from a loss of dopaminergic neurons in the substantia nigra in basal ganglia. The cause of these cell death is not well understood but is would be linked with an abnormal accumulation of alpha-synuclein protein. The standard diagnostic rely on motor tests that are positives when they reveal two out of three of the following symptoms: akynesia (difficulty to start a movement), rigidity and rest tremors. Unfortunately, those motor symptoms appear once 70% of dopaminergic neurons are destroyed. For this reason, the detection of Parkinson's disease at an earlier stage, before 70% of dopaminergic neurons are lost, represents one of the main goals of PD research. One method that has the potential to detect early changes in PD patients is the acoustic analyze of the voice. Indeed, people with Parkinson's disease have vocal impairments characterized by hypokinetic dysarthria. This includes perturbation in prosody, articulation and phonation. Their voice is more monotonous (with a diminution of intensity and pitch modulations). Speech flow is altered and patients make more dysfluencies. Consonant articulation is imprecise, particularly for occlusive consonants (p, t, k, b, d, g). Vowel articulation is also impaired: differences between vowels tend to decrease, which results in a reduction of the vowel surface area. As for phonation, pitch and intensity are unsteady (particularly during sustained vowel tasks), and timbre is hoarse. More details about those vocal impairments are given in [2]. One of the main interests of voice analysis in PD is that vocal impairments are present from the beginning of the disease and even several years before that a clinical diagnostic can be made [3]–[5]. Moreover automatic detection of PD based on voice's acoustic analysis reaches an accuracy of 95% [6]. As for specific detection of the early and middle stages of PD, the best accuracies are around 90% [5]–[7]. The speech tasks used for the vocal analyses in PD detection are most of the time: – Sustained vowel: vowel /a is the most used – Diadochokinesia (DDK) task: fast syllables repetition, usually syllable with occlusive consonant (like /pa/-/ta/-/ka/) This task reveals well the consonant articulation impairments – Reading and free speech: to analyze consonant and vowel articulation as well as prosody At the beginning all the studies in that field performed long time signal analyses. They extracted global features such as number of pauses, number of dysfluent words, standard deviation of pitch and of intensity. They also averaged local time perturbations, such as shimmer, jitter, voice onset time, signal to noise ratio, formants and vowel space area… in order to have one feature vector of one dimension per subject and per task. Once the features were extracted, they performed significance and redundancy tests to select a reduced set of features. Then they implemented the selected features in classifiers (support vector machine (SVM) was the most

Institut Mines-Télécom, Fondation Télécom, Institut Carnot Telecom & Société Numérique commonly used). Finally, most of the time, in order to have an estimation of the real accuracy, they validated their classifier with cross validation, because of the limited number of subjects. Leave one subject out (LOSO) and 10-fold cross validation were the two cross validation techniques the most used in this context. Most of those long term acoustic parameters require an accurate estimation of the fundamental frequency, which is a hard task, above all for pathological voices. On the other hand, other type of parameters were widely used in speaker and speech recognition: the mel-frequency cepstral coefficients (MFCC), associated with Gaussian Mixtures Models (GMM). MFCC are short-term parameters: they are calculated on a short time window (between 20ms to 40ms) every 10-20ms. They characterized the spectrum envelop on a MEL scale (which reflect the natural auditory perception). These features have the advantage to not depend on pitch estimation. For a decade, MFCCs have been used for the detection of different voice impairments, like dysphonia [8]–[10]. They were implemented in PD disease detection a few years later, the first time by Tsanas [11] and then by several other studies [12]–[15]. These authors extracted some statistics from the MFCCs (mean, SD and sometimes kurtosis and skewness) and added these features to the other classical features in their classifier. They calculated the statistics on the whole task when the task was acoustically steady (like sustained vowels) [11], [15] or on portions of the tasks that shared some acoustical characteristics, like voiced and unvoiced frames [12] or transitions frames (between voiced and unvoiced) [14]. In order to get information on frames that are acoustically very different (for example if we take all the frames from a reading task) we have to model more precisely the MFCCs distribution. One on the possible methods is to use multi-dimensional Gaussian Mixture Model (GMM) as an approximation of the multi-dimensional probability density function of MFCCs. For speech features it has been proved that a finite number of Gaussians (from 8 to 64) is sufficient to form a smooth approximation of the probability density function [16]. Bocklet and his coworkers [17] in 2013 used this method to model the MFCC distribution during different kind of tasks (sustained vowels, free speech, reading, DDK). They created one multi-dimensional GMM (with 128 Gaussians) per subject and per task with GMM-UBM modeling, and kept the means of the 128 Gaussians. Then they implemented those multi-dimensional feature vectors (one per task and per speaker) in a SVM classifier. With this method the best accuracy they obtained was 81% during a reading task. This method was also tested by [6] for three different languages during reading task (for which they obtained an accuracy around 80%) and during DDK task. In this last task the accuracies go from 70% to 87% depending on the language. In speaker recognition, separating men from women gives better classification results. Moreover it already has been proved that differentiation by gender increased performance in detection of some voice impairments, like laryngeal pathologies [18] when MFCCs are used. We decided to follow this direction and see how we can exploit the GMM modeling in order to distinguish early stages of PD from healthy controls. Unlike the previous studies which used GMM modeling in PD detection we created one GMM per class (not one per subject) and merely calculated the likelihood of the MFCCs distribution of test subjects against the 2 GMM models (one for PD class and one for control class). We wanted to see if we could have good classification performance for early stages of PD detection with this simple, automatic and fast method. To have the best performances possible we separated women from men. II.  DATA A.  Participants PD patients were recruited by neurologists, in the context of a longitudinal protocol (ICEBERG study) conducted at the hospital Pitié-Salpêtrière in collaboration with the clinical unit of the Institut Cerveau Moelle (ICM). Some patient’s spouses were recruited as control subjects. Other control subjects were recruited via an offer of the RISC (French information relay in cognitive science). Different inclusion criteria were applied depending of the participants’ status. The inclusion criteria for PD patients were: a diagnosis of idiopathic Parkinson’s disease during the last 4 years, no atypical Parkinsonian syndrome (like MSA, Lewy body dementia, PSP), no Parkinsonian syndrome due to neuroleptic or MPTP, a positive DatScan, corresponding to a possible, probable or definite idiopathic Parkinson disease according to UKPDSBB criteria [19]. The inclusion criteria for control subjects were: having a normal neurologic examination and being age-matched with the patients’ groups (between 40 and 70 years old). All those subjects (PD and controls) carried out several motor and cognitive tests, biological tests and medical imaging. For our study we recorded 74 French subjects among these participants. 40 were recently diagnosed with Parkinson’s disease (21 males and 19 females). 34 were healthy control subjects (14 males and 20 females). The mean age is 61.7 yr (+/- standard deviation (SD) 7.0 yr) for male PD, 62.4 yr (+/- SD 9.2 yr) for female PD, 54.9 yr (+/- SD 9.7 yr) for male controls and 54.8 yr (+/- SD 8.1 yr) for female controls. PD subjects are pharmacologically treated and the voice recordings occurred at different moments of the day: patients could be on the effect on their treatment (ON-state) or not (OFF-state). B.  Recording The participants were recorded once in a consultation box at the hospital Pitie Salpétrière in Paris. Their voice was recorded with a professional head mounted omnidirectional condenser microphone (Beyerdynamics Opus 55 mk ii) placed approximately 10 cm from the mouth. This microphone was connected to a professional sound card (Scarlett 2i2, Focusrite). Speech was sampled at 96000 Hz with 24 bits resolution, with a spectrum of [50Hz, 20kHz]. There was a

Institut Mines-Télécom, Fondation Télécom, Institut Carnot Telecom & Société Numérique preamplification in the external sound card connected to the head mounted microphone. C.  Speech Task The speech tasks were presented to the patients via an interface user on a computer programmed with Matlab. They had 28 short tasks to carry out. Each task lasted between 2 s and 1 min. The total of the procedure lasted 20 minutes. The tasks appeared with a random order. At the beginning of some tasks they could hear an audio example. If they did not do a task correctly, they could do it again. The speech tasks are presented in Table I. TABLE I TASKS WITH DESCRIPTION, OCCURENCE AND DURATION.
a.  The duration is the total duration (all the occurrences cumulated for one type of task) and is an average among the participants b.  DDK stands for diadochokinesia III.  METHODS: ACOUSTIC ANALYSIS AND CLASSIFICATION A.  MFCC extraction MFCCs were extracted using function melcepst of the Voicebox toolbox of matlab [20] with the default parameters. Twelve MFCCs were calculated in a time window of 20-ms length and with a 10-ms time shift, resulting in 12 MFCCs per frame with one frame every 10 ms. B.  Classification The goal of this classification was to be able to distinguish early PD patients from controls, knowing the gender of an unknown subject. We had speech data from early PD patients (males and females) and healthy controls (males and females) that we could use to build a statistical GMM model for each of the 4 groups. Knowing the gender of a new subject we wanted to see with which accuracy we could predict if she/he belonged to the PD or control side. During the development phase, also called training phase, we built 4 multi-dimensional GMM models. Means, SD and weights of the Gaussians (which formed the GMM) were estimated via an Expectation-Maximization algorithm. We could choose the number of Gaussian functions used to fit the distribution of the MFCCs. This number depended on the quantity of speech data we had per group (which depended on the number of subjects per group and the length of the task). Preliminary tests showed that 20 Gaussians seemed relevant for approximately 20 subjects per group for a 1-min task. To classify one test subject, we computed the log-likelihood of its MFCCs distributions against the PD and the control models (that corresponded to his or her gender). The model against which the test subject had the highest log-likelihood would determine his/her class. More precisely, we did not compute directly the log-likelihood of all the entire set of test speech data against the models, but we computed one log-likelihood, per frame of the data test, against the 2 models. Then we averaged the log-likelihoods on all the frames. This method guaranteed the likelihood to be independent of the frame number. C.  Validation As we the number of subjects was not large enough, we could not split subjects into two groups: one for the training and test and one for the validation. Therefore, to still have a correct estimation of the real accuracy of our classifier, we used a leave one subject out (LOSO) cross validation. We used all the subjects except one for the training of our models and we used the remaining subject for the test. We repeated the procedure so that each subject was tested once. Then we computed the classification accuracy (Acc_cv) based as the rate of well classified subjects. The real accuracy (the one we would obtain on an infinite number of new test subjects) is approximately equal to the cross validation accuracy we just computed. We estimated the density probability of the real accuracy by a Gaussian function with a mean equal to the cross validation accuracy, and a standard deviation given in (1). !"#=#√&''_')*(1-&''_'))/0 (1) with n being the number of subjects tested during the cross validation. Task Description Occurrence Durationa aaa sustained vowel /a as long and steady as possible without breathing 2 20 s DDKb repetition of syllables as fast and steady as possible without breathing (pa, pu, ku, pupa, paku, pataka, badaga, patiku, pabiku, padiku) all : 1 except pataka : 2 2 min reading reading of 2 short sentences (one question and one exclamative sentence) , 1 short text and 1 dialog. 1 each 1 min free speech telling about one's day 1 1 min glissando sustained vowel /a while varing the pitch: from deep to high then to deep again , without breathing 2 15 s repetition repetition of 4 short sentences (2 questions and 2 exclamative sentences) pre-recorded 1 each 10 s rythm saying some syllables (pa, ku or pa ku) following a rythm previously heard in a recording (1 syllable per sec) during 30s pa : 1 ku: 1 pa ku : 2 2 min

Institut Mines-Télécom, Fondation Télécom, Institut Carnot Telecom & Société Numérique Actually this provides a good estimation of the real accuracy when there are enough tests (>30) and when the tests are iid (independent and identically distributed) which is not completely the case in a cross validation (the tests are not completely independent). But for stable inducers we can use those formula for LOSO cross validation [21]. In order to have an estimation of the true positives and the true negative we also computed the sensitivity (rate of PD correctly classified) and the specificity (rate of controls correctly classified) of our classifier. The estimation of the real sensitivity and specificity values follow the same rule as for the accuracy. D.  Statistical tests In order to compare the different performances obtained with the LOSO we calculated several statistical tests. As our number of subjects tested in each classification test was higher than 30 we used a Welch's t-test to compare the accuracies of different populations (e.g. men vs women). We used a paired t-test to compare the accuracies during different tasks carried out by a same population (e.g. only men or only women). P value was computed each time and gave us information about the significances of the differences. We fixed the risk at 5%. IV.  RESULTS For now we performed the acoustic analyses and the LOSO cross validation for 4 of our tasks (the most common tasks): the sustained phonation vowel task, the DDK task, the reading task and the free speech task. As mentioned previously, we tested men and women separately, because of the influence of the gender on the MFCC, and together (for the comparison). To avoid a gender bias we removed (at random) 2 male PD and 6 female controls so that the PD model and the control model had the same amount of males and females each. We tested our classifier with various numbers of Gaussian mixtures for our models and we had confirmation that 20 Gaussians was the best number for our tasks. The results from the LOSO cross validation calculated with 20 Gaussians for our models are presented in Table II. Our classifier, based on a MFCC distribution and GMM modeling of healthy and early PD patients, showed that the reading task seemed to be the most relevant for our classification method (as compared to the three classical others: sustained vowel, DDK and free speech). The reading task's score was significantly better than the sustained vowel one (for both men (p=0.0004) and women (p=0.03)). Men obtained also significantly higher classification scores during reading task than free speech tasks (p=0.03). For the reading task the accuracy of our classifier for male controls vs male PD was 91.4 %, and 79.5% for female controls vs female PD. TABLE II CLASSIFICATION PERFORMANCE OBTAINED WITH LEAVE ONE SUBJECT OUT CROSS VALIDATION.
Acc stands for accuracy, Se for Sensitivity, Sp for Specificity, SD for Standard Deviation, PD for Parkinson's Disease, aaa for sustained vowel "a", DDK for diadochokinesia. Results are presented in %. When grouping male and female together, Acc was 75.8% . Accuracy was thus higher for males alone than females alone (although differences were not significant (p =0.14) during the reading task). The score of males alone was also higher than males and females together, and the difference was significant (p=0.03). It seems better for men to classify them separately, but it is not so obvious for women. In order to answer this question, we calculated the accuracy when classifying females between a PD group (men and women mixed) and a control group (men and women mixed). The accuracy was 78.8% (+/- SD 7.1%), which is in the same range as the one obtained with female alone (no significant differences: p=0.95 for the reading task). By contrast, the difference between the accuracy obtained for men classified with male models and with mixed models was significant (p=0.04 for the reading task). Therefore classifying males and females between PD group and control group is more powerful for men classification when we separate them. As the sustained phonation task was shorter than the reading task we tested if its low classification performance could be explained by the short duration. Indeed generally speaking the more speech data we have, the better we can train the GMM model and the higher is the performance of the classification. To check the influence of the duration of the task on the performance, we performed a classification on a short part of the reading (we just kept the 2 short sentences). This new task lasted 4 sec. After performing a LOSO cross validation we obtained an accuracy of 88.6% for males and 74.4 % for females. These score are a little lower than the scores for the entire reading task, even though the difference is not significant (p=0.57 for men and p=0.42 for women). This slight difference is coherent with the positive role of the duration in the performance of the classifier. Nevertheless this shorter reading task stay better than the other tasks, and well better than the sustained vowel task, confirming the relevancy of a reading task for our classifier. Therefore we can conclude Tasks PD men vs Control men SD PD women vs Control women SD PD vs Controls SD Acc aaa Se 20s Sp 60.0 8.3 71.4 9.9 42.9 13.2 59.0 7.9 42.1 11.3 75.0 9.7 48.5 6.2 55.3 8.1 39.3 9.2 Acc DDK Se 2min Sp 82.9 6.4 81.0 8.6 85.7 9.4 66.7 7.5 57.9 11.3 75.0 9.7 74.2 5.4 71.1 7.4 78.6 7.8 free Acc speech Se 1min Sp 74.3 7.4 85.7 7.6 57.1 13.2 71.8 7.2 78.9 9.4 65.0 10.7 65.2 5.9 71.1 7.4 57.1 9.4 Acc reading Se 1min Sp 91.4 4.7 95.2 4.6 85.7 9.4 79.5 6.5 78.9 9.4 80.0 8.9 75.8 5.3 84.2 5.9 64.3 9.1 reading Acc (extract) Se 4s Sp 88.6 5.4 95.2 4.6 78.6 11.0 74.4 7.0 73.7 10.1 75.0 9.7 78.8 5.0 78.9 6.6 78.6 7.8

Institut Mines-Télécom, Fondation Télécom, Institut Carnot Telecom & Société Numérique that the short duration of the sustained vowel task was not the main cause of its low score. V.  DISCUSSION The MFCC distribution method we used (method widely used for speaker recognition), and which was never used this way (as far as we know) in early PD detection, showed results in line with the best performances found in the literature : around 90% for the detection of early stages of PD for men [6]. In this study [6] the performance obtained with just MFCC (but with another method) reaches 85% for early PD detection. According to our study, the MFCC distribution method seemed to be specifically relevant for the reading task (Acc=91% for men), and especially for the reading of short questions or exclamation sentences (Acc=88% for only 4 sec of speech data per men). The fact that the reading task had better performance than free speech with equal duration may be explained by the fact that in the reading task, the variability depends only on the speakers' voice and on the effect of PD on their voice, whereas in the free speech task the variability depends also on the content (the sentences and words chosen). The sustained vowel task seemed particularly inappropriate for the MFCC distribution method (Acc= 60% for men, that is just a little better than random would do). Our study revealed that splitting males and females increased the classification performance, which is line with what is known in speaker recognition. Actually it increased particularly male classification performances. The fact that we obtain lower performance for females than for males may be explained by the higher average pitch of females that renders cepstral analyses less accurate. Moreover it is also the case for the detection of other voice disorders such as laryngeal disorders [18]. We may have obtained even better performances if the PD subjects were off medication because the PD medications improve vocal perturbations [22]. In the future we will carry out additional analyses that may improve our results. We will perform the acoustic analyses on the three other speech tasks. We will add new subjects (PD and controls). We will add delta MFCCs, because studies showed delta MFCC could carry additional information for the assessment of parkinsonian voices perturbation [23]. Finally we will combined this classification method, based on GMM and short-term MFCC analyze, with more classical methods using global features that are discriminant in the detection of stages of PD (cf annex 2 of [2]). VI.  CONCLUSION An automatic and simple methodology, based on the one used for speaker recognition, was presented to detect early stages of Parkinson's disease. We performed an automatic analysis of the recordings of early PD patients and healthy controls during 4 types of tasks: sustained vowel, DDK, free speech and reading. The analysis methodology consisted to extract 12 MFCCs every 10 ms for each task, and to fit their distribution with 2 multi-dimensional GMMs: one for PD model and one for control model. The log-likelihood of the test subjects MFCC distributions against those 2 models allowed a 2-classes classification between PD group and control group. The results obtained with a LOSO cross validation seems to indicate that the reading task is particularly appropriate with this methodology. The accuracy obtained (91% for men) is in line with the best performance that we can find in the literature concerning early PD detection. The mix of this analysis method and other more classical method in PD detection may improve performance for women and for the other vocal tasks. REFERENCES [1] L. M. De Lau and M. M. Breteler, ‘Epidemiology of Parkinson’s disease’, Lancet Neurol., vol. 5, no. 6, pp. 525–535, 2006. [2] L. Jeancolas, D. Petrovska-Delacrétaz, S. Lehéricy, H. Benali, and B.-E. Benkelfat, ‘L’analyse de la voix comme outil de diagnostic précoce de la maladie de Parkinson  : état de l’art’, in CORESA 2016  : 18e {’e}dition COmpressions et REpr{’e}sentation des Signaux Audiovisuels, Nancy, 2016, pp. 113–121. [3] B. Harel, M. Cannizzaro, and P. J. Snyder, ‘Variability in fundamental frequency during speech in prodromal and incipient Parkinson’s disease: A longitudinal case study’, Brain Cogn., vol. 56, no. 1, pp. 24–29, Oct. 2004. [4] R. B. Postuma, A. E. Lang, J. F. Gagnon, A. Pelletier, and J. Y. Montplaisir, ‘How does parkinsonism start? Prodromal parkinsonism motor changes in idiopathic REM sleep behaviour disorder’, Brain, vol. 135, no. 6, pp. 1860–1870, Jun. 2012. [5] J. Rusz et al., ‘Quantitative assessment of motor speech abnormalities in idiopathic rapid eye movement sleep behaviour disorder’, Sleep Med., Sep. 2015. [6] J. R. Orozco-Arroyave et al., ‘Automatic detection of Parkinson’s disease in running speech spoken in three different languages’, J. Acoust. Soc. Am., vol. 139, no. 1, pp. 481–500, 2016. [7] M. Novotny, J. Rusz, R. Cmejla, and E. Ruzicka, ‘Automatic Evaluation of Articulatory Disorders in Parkinson #x2019;s Disease’, IEEEACM Trans. Audio Speech Lang. Process., vol. 22, no. 9, pp. 1366–1378, Sep. 2014. [8] A. A. Dibazar, S. Narayanan, and T. W. Berger, ‘Feature analysis for automatic detection of pathological speech’, in Proceedings of the Second Joint 24th Annual Conference and the Annual Fall Meeting of the Biomedical Engineering Society] [Engineering in Medicine and Biology, 2002, vol. 1, pp. 182–183 vol.1. [9] J. I. Godino-Llorente and P. Gomez-Vilda, ‘Automatic Detection of Voice Impairments by Means of Short-Term Cepstral Parameters and Neural Network Based Detectors’, IEEE Trans. Biomed. Eng., vol. 51, no. 2, pp. 380–384, Feb. 2004. [10] N. Malyska, T. F. Quatieri, and D. Sturim, ‘Automatic dysphonia recognition using biologically-inspired amplitude-modulation features’, in Acoustics, Speech, and Signal Processing, 2005. Proceedings.(ICASSP’05). IEEE International Conference on, 2005, vol. 1, p. I–873. [11] A. Tsanas, M. A. Little, P. E. McSharry, J. Spielman, and L. O. Ramig, ‘Novel Speech Signal Processing Algorithms for High-Accuracy

Institut Mines-Télécom, Fondation Télécom, Institut Carnot Telecom & Société Numérique Classification of Parkinson's Disease’, IEEE Trans. Biomed. Eng., vol. 59, no. 5, pp. 1264–1271, May 2012. [12] J. R. Orozco-Arroyave et al., ‘Automatic detection of parkinson’s disease from words uttered in three different languages.’, in INTERSPEECH, 2014, pp. 1573–1577. [13] J. R. Orozco-Arroyave et al., ‘Characterization Methods for the Detection of Multiple Voice Disorders: Neurological, Functional, and Laryngeal Diseases’, IEEE J. Biomed. Health Inform., vol. 19, no. 6, pp. 1820–1828, Nov. 2015. [14] J. R. Orozco-Arroyave et al., ‘Towards an automatic monitoring of the neurological state of Parkinson’s patients from speech’, 2016, pp. 6490–6494. [15] D. Hemmerling, J. R. Orozco-Arroyave, A. Skalski, J. Gajda, and E. Nöth, ‘Automatic Detection of Parkinson’s Disease Based on Modulated Vowels’, 2016, pp. 1190–1194. [16] T. F. Quatieri, Discrete-Time Speech Signal Processing: Principles and Practice, 1 edition. Upper Saddle River, NJ: Prentice Hall, 2001. [17] T. Bocklet, S. Steidl, E. Nöth, and S. Skodda, ‘Automatic evaluation of parkinson’s speech-acoustic, prosodic and voice related cues.’, in Interspeech, 2013, pp. 1149–1153. [18] R. Fraile, N. Sáenz-Lechón, J. I. Godino-Llorente, V. Osma-Ruiz, and C. Fredouille, ‘Automatic Detection of Laryngeal Pathologies in Records of Sustained Vowels by Means of Mel-Frequency Cepstral Coefficient Parameters and Differentiation of Patients by Sex’, Folia Phoniatr. Logop., vol. 61, no. 3, pp. 146–152, Jul. 2009. [19] A. J. Hughes, S. E. Daniel, L. Kilford, and A. J. Lees, ‘Accuracy of clinical diagnosis of idiopathic Parkinson’s disease: a clinico-pathological study of 100 cases.’, J. Neurol. Neurosurg. Psychiatry, vol. 55, no. 3, pp. 181–184, 1992. [20] ‘VOICEBOX’. [Online]. Available: http://www.ee.ic.ac.uk/hp/staff/dmb/voicebox/voicebox.html. [Accessed: 20-Mar-2017]. [21] R. Kohavi, ‘A study of cross-validation and bootstrap for accuracy estimation and model selection’, in Ijcai, 1995, vol. 14, pp. 1137–1145. [22] J. Rusz et al., ‘Evaluation of speech impairment in early stages of Parkinson’s disease: a prospective study with the role of pharmacotherapy’, J. Neural Transm., vol. 120, no. 2, pp. 319–329, Feb. 2013. [23] A. Tsanas, M. A. Little, P. E. McSharry, and L. O. Ramig, ‘Nonlinear speech analysis algorithms mapped to a standard metric achieve clinically useful quantification of average Parkinson’s disease symptom severity’, J. R. Soc. Interface, vol. 8, no. 59, pp. 842–855, Jun. 2011.