Comparing Parkinson’s Disease Dysarthria and Aging Speech using

Articulation Kinematics

A. Gómez-Rodellar

, D. Palacios-Alonso

, J. Mekyska

, A. Álvarez-Marquina

and

P. Gómez-Vilda

Neuromorphic Speech Processing Lab, Center for Biomedical Technology, Universidad Politécnica de Madrid,

Campus de Montegancedo, 28223 Pozuelo de Alarcón, Madrid, Spain

Escuela Técnica Superior de Ingeniería Informática, Universidad Rey Juan Carlos, Campus de Móstoles, Tulipán,

s/n, 28933 Móstoles, Madrid, Spain

Department of Telecommunications, Brno University of Technology, Technicka 10, 61600 Brno, Czech Republic

Keywords: Parkinson’s Disease, Neuromorphic Speech Processing, Neurotechnology, Aging Speech, e-Health.

Abstract: Speech is being considered a pervasive and costless means to detect and monitor neurodegenerative disease

progression. Many different approaches have been reported to differentiate normative subject speech from

neurodegenerative patient speech. Most of them are focussed on statistical pattern recognition approaches to

improve detection results on a baseline, considering only patient speech and normative controls. The

definition of a normative control is not well established in itself, usually being subjects free of any pathology

aligned in the same age range as patients. But one question which is not taken into account is the effects of

aging in healthy controls, as usually neurodegenerative diseases may include mostly patients affected by

certain effects, as dysphonia or dysarthria, as a consequence of aging. The present research introduces a

methodology based on information theory to compare the effects produced by aging dysarthria with those due

to Parkinson’s Disease, using the statistical distribution of speech articulation kinematics as a marker. On the

one hand, it may be concluded that articulation kinematics is substantially different for PD and HC with

respect to normative subjects. On the other hand, this does not seem to be the case between PD and HC

subjects, as these subsets may share some dysarthric features which may be contributed more by aging than

by neuromotor degeneration. This differentiation problem needs to be evaluated as well in the case of

phonation features, otherwise there will not be full guarantee in using phonation features to assess neuromotor

degeneration. In this sense new methodologies have to be designed to distinguish neurodegenerative from

aging speech granting better guarantees.

1 INTRODUCTION

Neurodegenerative diseases have a clear effect on

speech, both in phonation, articulation, prosody and

fluency. Parkinson’s Disease (PD) is among the most

prevalent neurodegenerative diseases, affecting

around 5 million people over age 50 in the 15 world

most populated countries in 2005, doubling by 2030

(Dorsey et al., 2007). Typical symptoms associated to

PD are bradykinesia, rigidity, freezing of gait, frozen

facial mask (hypomimia), postural sway, and distal

limb resting tremor, among others (Dauer and

Przedborski, 2003, Jankovic, 2008; Sapir, 2014;

Anizah et al., 2018). It is well known that speech is

strongly related to axial symptoms (Gobermann,

2005; Cantiniaux et al., 2010; Ricciardi et al. 2016).

Phonation, articulation, prosody and fluency are

speech characteristics strongly affected by PD.

Phonation symptoms (musculus vocalis hypotonia),

vocal fold unbalance and tremor (altered neuromotor

feedback) are some ways in which the

neurodegeneration manifests. Articulatory instability

is observed mainly as reduced vowel space and vowel

centralization distortion (Sapir et al., 2010).

Dysprosody and dysfluency are also common

symptoms having received attention (Goberman,

Blomgren and Metzger, 2010; Martens, et al., 2015).

A view of the most comprehensive studies in the field

can be found in Tsanas et al. (2010), Rusz et al.

(2013), Mekyska et al. (2015), and Brabenec et al.

(2017). The objective of this study is to compare

articulation in PD patients and aging healthy controls

against a normative population, using kinematic

features estimated from formants, relying on

Information Theory to determine if the steady jaw

control necessary to maintain a vowel in its precise

articulation place is similarly affected by aging voice

Gómez-Rodellar, A., Palacios-Alonso, D., Mekyska, J., Álvarez-Marquina, A. and Gómez-Vilda, P.

Comparing Parkinson’s Disease Dysarthria and Aging Speech using Articulation Kinematics.

DOI: 10.5220/0007355700520061

In Proceedings of the 12th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2019), pages 52-61

ISBN: 978-989-758-353-7

than by PD or not and quantify in thus wat the

divergence of pathological and aging articulation

with respect to normative subjects, to pinpoint

differences and similarities. Recent studies have

shown that a relationship can be established between

formant-based articulation features and jaw-tongue

kinematic activity. This relationship allows to

estimate the jaw-tongue kinematics from formant

dynamics. The mutual information contents from

probability density functions of jaw-tongue kinematic

activity estimated from formant-based articulation

features may be used as dysarthria markers when

comparing PD speech with normative speech

(Gómez, P. et al., 2018). The neuromotor character of

these markers has been validated by facial surface

electromyography and accelerometry (sEMG and

3DAcc), as shown in (Gomez, A. et al., 2018).

Building on this relationship, the purpose of the

present work is to explore if these dysarthria markers

are affected differently in presbyphonic dysarthria

(characteristic of aging speech) than in PD dysarthria.

The structure of the paper is as follows: Section 2 is

devoted to explain the foundations of the jaw-tongue

kinematics related with speech articulation

modelling, the distributions describing the statistical

behaviour of the kinematic variables associated with

articulation, and the validation of the kinematic

correlates. Section 3 describes the databases of

normative, healthy controls and patients, used in the

differentiation experiments, and the mutual

information estimation methods. Section 4 gives a

complete description and discussion of the results

produced by the differentiation experiments, both as

tables and as figures representing the proximity or

distance of each sample to the average of the

reference sets (nominally, normative and healthy

controls). Finally, section 5 is devoted to highlight the

conclusions derived from the presented results.

2 ARTICULATION KINEMATICS

2.1 Jaw-Tongue Biomechanical Model

Speech articulation depends on the position and shape

of vocal tract structures, such as the jaw, tongue, lips

and velo-pharynx, among others (Buchaillard, Perrier

and Payan, 2009). These structures are controlled by

different muscles, which are activated by neuromotor

pathways from cranial nerves (Jürgens 2002). The

acoustical characteristics of speech sounds depends

on the positions of these structures and on their

dynamic displacement. In the present paper, the role

of the jaw-tongue system, as depicted in Figure 1 will

be studied when affected by neuromotor degeneration

induced by PD. The jaw-tongue biomechanical

system is considered to be a third-order lever with

lumped mass load concentrated in the reference point

rJT

, y

} (Hannam et al., 2008). Harmonic

oscillation {Δx

, Δy

} around the fulcrum (F:

attachment to the skull) is assumed under forces

acting on this system. A very relevant kinematic

correlate of the jaw-tongue neuromotor activity is the

Absolute Kinematic Velocity (AKV) of the reference

point PrJT:







































(1)

The statistical distribution of the AKV will

contribute valuable information in characterizing

unstable articulation, as explained in the sequel.

Figure 1: Jaw-Tongue Model. F: Fulcrum; T: Tongue; J:

Jaw bone; H: Hyoid bone; f

: stylo-glossus force; f

masseter force; f

: glosso-intrinsic forces; f

: genio-hyoid

force; f

: gravity; X

: accelerometer normal and

tangential; Δx

, Δy

: horizontal and vertical displacements

of the reference point (P

rJT

) in the sagittal plane.

2.2 A Kinematic Articulation Correlate

The methodology of this research is based on

representing speech articulation kinematics

(positions, speeds, forces and accelerations) by means

of acoustically-derived information (speech

formants; Dromey, Jang and Hollis, 2013). An

important question on the use of kinematic features

derived from acoustic correlates (the first and second

formants: F

and F

) is to which extent formant

dynamics can be related to articulation kinematics

(positions and velocities of the jaw-tongue centre of

masses). The assessment of the AKV as a reliable

kinematic correlate of articulation is carried on the

multi-signal recording framework described in

Comparing Parkinson’s Disease Dysarthria and Aging Speech using Articulation Kinematics

Figure 2. The experimental validation of using

acoustic information (formant-based dynamics) to

represent articulation kinematics was based on a

diadochokinetic exercise, consisting in the fast and

continuous repetition of the diphthong [aj:], at a rate

of 2-3 repetitions per second. Inverse adaptive

filtering was used to estimate the vocal tract transfer

function from running speech in real time (Deller,

Proakis and Hansen, 1993). F

and F

are evaluated

from the vocal tract transfer function obtained from

inverse filtering. Surface electromyography on the

masseter (sEMG) and three-channel accelerometry

(3DAcc) were recorded synchronously with speech.

Sampling rates of sEMG and 3DAcc were equalized

to 500 Hz, as well as formant estimates.

Figure 2: Recording set-up for Signal acquisition of speech,

accelerometry and surface electromyography (sEMG).

The validation of formant dynamics to represent

kinematic variables was based on linear regression

according to the following relational chain: surface

electromyography (sEMG) is related to the force on

the masseter (f

), which on its turn is related to

vertical acceleration (yAcc), resulting in vertical

displacement (Δy

), changing the vertical articulation

position, which induces changes in the first two

formants (ΔF

, ΔF

). The results of regression studies

among the different dynamic variables are given in

Table 1.

Table 1: Regression results for the diadochokinetic

validation exercise. r: correlation coefficient; p: p-value; S:

Spearman’s coefficient; P: Pearson’s coefficient.

Correlation

r (S)

p (S)

r (P)

p (P)

Δy

vs f

0.83

<0.001

0.81

<0.001

ΔF

vs Δy

-0.89

<0.001

-0.89

<0.001

ΔF

vs Δy

0.78

<0.001

0.79

<0.001

The correlation between the masseter force

estimate from sEMG (f

) and the vertical

displacement of the reference point (Δy

) is high and

statistically relevant (0.83/0.81), showing that a

strong relationship exists between neuromotor

activity and movement, as expected. The correlation

between vertical displacement (Δy

) and formant

changes are also high and relevant, stronger and

counter-related with respect to ΔF

(-0.89), than with

respect to ΔF

(0.78/0.79). These results are aligned

with the relationship between the variable controlling

the phonation opening (Δy

) and the variation of the

first formant (ΔF

). Once the relationship between

kinematics and acoustics has been established and

validated, the displacement of the reference point of

the jaw-tongue system when observed over time

could be described from an estimate of (1) as:































































(2)

where F

and F

are the first two formants, and B

, B

and B

are quadratic scaling factors relating

movement and acoustics (Gómez, A., et al., 2018).

The distribution of the AKV values as a probability

density function (AKV pdf) gives a full statistical

description of the jaw-tongue kinematics, and of the

kinetic energy which is involved in speech

production. The shape of the AKV probability density

function will be that of a χ

(Chi-square) distribution

with two degrees of freedom, which is typically

associated with thermodynamic processes, justifying

the use of the term “low articulation temperature”

associated to hypokinetic dysarthria, an example of

these distributions shown in Figure 3.

Figure 3: Two ideal probability density functions associated

to the AKV in terms of their respective “temperatures”. The

model distribution would be associated to the AKV pdf of

a maintained vowel from an idealized normative speaker,

whereas the target distribution is the typical behaviour of a

vowel from a PD patient.

It must be considered that the behaviour of the

AKV pdf is quite different according to the kinematic

study being carried on. When steady vowels are

BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing

produced, as in single maintained vowel exercises, it

is expected that a normative speaker would keep a

highly stable jaw-tongue position (low temperature)

with most of the absolute velocities under a given

value (dash-diamond curve), whereas the PD patient

will produce unstable oscillations of the articulation

point (high temperature) extending along the

horizontal axis (full-bullet curve). The situation in

running speech, where wide oscillations of the

reference point will be expected is to be the opposite:

the normative speaker will produce wider and faster

oscillations (higher temperature) than the PD patient

affected by hypokinetic dysarthria (lower

temperature). This fact points to a complete different

strategy in testing sustained vowels than in running

speech or dyadochokinetic exercises.

3 MATERIALS AND METHODS

3.1 Patient Data Sets

In the present study the articulation stability in

maintained vowels has been used to assess the

capability of these tests in differentiating the

behaviour of PD patients from healthy controls within

the same age range, when compared with a normative

reference set considered the golden rule in maintained

vowel phonation. For such, a three band study has

been conducted in terms of the mutual information

between AKV pdf’s from PD patients and paired

healthy controls, and with respect to normative

speakers, using correlation results from confronting

the three sets of speakers among themselves in terms

of Jensen-Shannon Distance (JSD). Estimates of the

AKV pdf have been used to evaluate the JSD between

two different distributions. Vowel utterances [a:, i:,

u:] from 8 male and 8 female PD patients randomly

selected from male and female databases within an

age range of 66.3±8.6 and 69±7.7 years (respectively)

have been processed and statistically modelled to

produce a PD database (MPD from male subjects, and

FPD for female ones). Similar vowel utterances from

another set of 8 male and 8 female control subjects

randomly selected from male and female databases

within an age of 65.6±8.9 and 61.8±9.1 years old

(respectively) have also been processed and

statistically modelled to produce a healthy control

database (MHC from male subjects and FHC from

female ones). Recordings were taken at 16 kHz and

16 bits. The database (PARCZ) was collected at St.

Anne’s University Hospital in Brno (Czech

Republic), including also demographic and clinical

information from each patient as gender, age, time

since first diagnosis, scores of the Unified

Parkinson’s Disease Rating Scale, part III (UPDRS-

III: motor examination), and part IV (UPDRS-IV:

complications of therapy), freezing of gait

questionnaire (FOG-Q), non-motor symptoms scale

(NMSS), REM sleep disorders (RBDSQ), mini-

mental state examination (MMSE), Addenbrooke’s

cognitive evaluation revised (ACE-R), Beck

depression inventory (BDI), faciokinesis and

phonorespiratory competence. All patients signed an

informed consent form that was approved by the local

ethics committee. The speakers extracted from the

PARCZ database are PD patients with code P1xxx

(females) and P2xxx (males), and paired healthy

controls with code K1xxx (females) and K2xxx

(males), as described in Table 2.

Table 2: PD patient and HC subject set lists (PD: PD patient

subject; HC: healthy control subject; UPDRS: Evaluation

according to UPDRS-III scale).

Code

Gender

Age

Cond

UPDRS

K1003-aiu

K1004-aiu

K1005-aiu

K1006-aiu

K1007-aiu

K1012-aiu

K1017-aiu

K1018-aiu

K2001-aiu

K2002-aiu

K2009-aiu

K2010-aiu

K2011-aiu

K2013-aiu

K2014-aiu

K2015-aiu

P1006-aiu

P1007-aiu

P1008-aiu

P1020-aiu

P1021-aiu

P1022-aiu

P1025-aiu

P1026-aiu

P2005-aiu

P2009-aiu

P2010-aiu

P2012-aiu

P2017-aiu

P2018-aiu

P2019-aiu

P2023-aiu

Comparing Parkinson’s Disease Dysarthria and Aging Speech using Articulation Kinematics

Finally, 8 male and 8 female subjects have been

randomly selected from a normative database

recorded at Hospital Gregorio Marañón, of Madrid,

Spain, within an age range of 3412.95 and 3713.37

(years) respectively. The list of subjects is given in

table 3.

Table 3: Normative subject set (NS).

Code

Gender/Age

Code

Gender/Age

N1004-aiu

M/23

N1105-aiu

F/43

N1005-aiu

M/21

N1108-aiu

F/22

N1008-aiu

M/45

N1112-aiu

F/20

N1009-aiu

M/33

N1116-aiu

F/45

N1011-aiu

M/49

N1117-aiu

F/25

N1018-aiu

M/29

N1120-aiu

F/33

N1020-aiu

M/35

N1121-aiu

F/57

N1026-aiu

M/39

N1125-aiu

F/38

3.2 Data Processing

The methodology proposed in the present study is

based on the mutual information between two given

probability density functions, p(x) and q(x) estimated

as a Jensen-Shannon Divergence (Endres and

Schindelin, 2003):





































(3)

where DKL is a modified version of Kulback-

Leibler’s Divergence (Salicrú et al., 1994; Georgiou

and Lindquist, 2003) expressed as:





 



















(4)

and m(x) is the average of p(x) and q(x). In the present

case, the probability functions p(x) and q(x) are

defined in the positive part of the real axis (x0).

Jensen-Shannon’s Divergence is symmetrical with

respect to p(x) and q(x), and it is normalized to the

interval [0, 1], a feature which is very helpful in

implementing clustering and classification. The

following procedure is used to estimate the JSD’s

between the PD set, the HC set and the NS set using

their AKV pdf’s:

• Recordings of the vowel set [a:, i:, u:] were

downsampled to 8 kHz.

• The vocal tract transfer function of the speech

segment was evaluated by an 8-pole adaptive

inverse lattice-ladder filter (Deller, Proakis and

Hansen, 1993) with a low-memory adaptive step

to grasp fine time variations. A complete

description of the adaptive filtering details can be

found in Gómez et al. (2009).

• The first two formants were estimated by

evaluating the maxima and slenderness of the

vocal tract transfer function (LP spectrogram).

The formant estimation resolution used was 2

Hz. Formants were estimated every 2 ms.

• The derivatives of the first two formants were

used to estimate the AKV following (2).

• The probability density function of the AKV was

estimated from the normalized histogram of

counts on the definition interval of the AKV (in

this case 0≤|v

|≤45 cm/s).

• The histograms were used to estimate probability

density functions by Kolmogorov-Smirnov

approximations (Webb, 2003).

• The average pdf for each subset was estimated. It

may be shown that the average of a set of pdf’s

shares the same properties of individual pdf’s.

Six average pdf’s were estimated: avMNS,

avFNS, avMHC, avFHC, avMPD and avFPD,

for the respective male and female normative,

controls and PD subsets.

• The Jensen-Shannon Divergence between each

patient’s histogram-derived distribution vs that

of the control subject were estimated as by (3)

4 RESULTS AND DISCUSSION

JSD’s between avMNS, avMHC and avMPD on one

side, and avFNS, avFHC and avFPD were estimated.

The divergences of the MPD vs MNS averages are

shown in table 4.

Table 4: JSD between male and female subset averages.

Datasets

JSD

avMPD vs avMNS

0.226

avMHC vs avMNS

0.244

avMPD vs avMHC

0.083

avFPD vs avFNS

0.311

avFHC vs avFNS

0.329

avFPD vs avFHC

0.092

The top template in Figure 4 shows the actual

appearance of the PD male sample AKV pdf’s in

dash-red, whereas the NS male AKV pdf’s are given

in full-blue. It may be easily seen that the NS set is

more concentrated towards the vertical axis, most of

the distributions having decayed on the interval

between 5-10 cm.s

-1

, whereas the PD set is more

spread over, with some activity still seen between 10-

30 cm.s

-1

and even beyond. The upper-right legend

gives the codes of the speaker samples included in the

tests. The bottom template in Figure 4 gives the

BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing

average pdf’s of the samples in the top template. The

different spreads of both average pdf’s may be clearly

seen now. The central legend gives the JSD between

both sets, as well as the results of the p-values after

Student’s, Kolmogorov-Smirnov’s, and Mann-

Whitney’s tests on Target and Model sets rejecting

the null hypothesis of equal means. The AKV pdf’s

from samples in the male and female subsets have

been obtained to be compared against the respective

normative and control subset averages (avMNS,

avMHC, for male samples, and avFNS and avFHC,

for female samples). Table 5 gives the JSD for each

sample. It may be seen that the divergence of the male

PD and HC average pdf’s with respect to the

normative one (MPD vs MNS and MHC vs MNS) is

quite similar and larger than when comparing PD and

HC (MPD vs MHC).

Figure 4: Top: AKV pdf’s of the PD male subset (dash-red lines) and the NS male subset (full-blue lines). The normative

subset is confined to lower absolute values than the PD subset. Bottom: Averages of normative (dash-blue) and PD (full-

red) pdf’s, showing the same behaviour. The JSD divergence between avMPD and av1MNS (0.226), and the p-values

rejecting the equal mean hypothesis by t-Student, Kolmogorov-Smirnov and Mann-Whitney tests are given in the middle.

Comparing Parkinson’s Disease Dysarthria and Aging Speech using Articulation Kinematics

Table 5: JSD’s between PD, healthy control and normative sets with respect to normative and control averages.

MPD

avMNS

avMHC

MHC

avMNS

avMHC

MNS

avMNS

P2005-aiu

0.30570

0.18871

K2001-aiu

0.24403

0.16737

N1004-aiu

0.074096

P2009-aiu

0.23901

0.14816

K2002-aiu

0.28961

0.15321

N1005-aiu

0.103380

P2010-aiu

0.13407

0.18628

K2009-aiu

0.32015

0.17684

N1008-aiu

0.124710

P2012-aiu

0.20950

0.21612

K2010-aiu

0.19482

0.16206

N1009-aiu

0.075808

P2017-aiu

0.40723

0.27555

K2011-aiu

0.31634

0.19989

N1011-aiu

0.153810

P2018-aiu

0.32686

0.18397

K2013-aiu

0.35297

0.17601

N1018-aiu

0.086608

P2019-aiu

0.23515

0.14976

K2014-aiu

0.23895

0.14632

N1020-aiu

0.088998

P2023-aiu

0.24250

0.23716

K2015-aiu

0.21416

0.17490

N1026-aiu

0.090407

FPD

avFNS

avFHC

FHC

avFNS

avFHC

FNS

avFNS

P1006-aiu

0.38868

0.18533

K1003-aiu

0.37033

0.17377

N1105-aiu

0.072714

P1007-aiu

0.25251

0.19843

K1004-aiu

0.34321

0.16863

N1108-aiu

0.093113

P1008-aiu

0.49834

0.27077

K1005-aiu

0.23298

0.19368

N1112-aiu

0.115570

P1020-aiu

0.21117

0.22827

K1006-aiu

0.47693

0.24286

N1116-aiu

0.120860

P1021-aiu

0.37974

0.17551

K1007-aiu

0.39329

0.15067

N1117-aiu

0.110290

P1022-aiu

0.28630

0.19981

K1012-aiu

0.26424

0.18785

N1120-aiu

0.104500

P1025-aiu

0.34688

0.17790

K1017-aiu

0.29882

0.17209

N1121-aiu

0.165080

P1026-aiu

0.29915

0.20053

K1018-aiu

0.44165

0.21344

N1125-aiu

0.076065

Figure 5: Distribution of each male sample with respect to the male HC and NS averages. a) Bi-plot in terms of JSD respect

to the mormative and healthy control sets (males). b) Geometrical distribution with respect to the centroids avHC and avNH.

c) and d) Similar representations for the female sets. Red squares: PD samples. Blue diamonds: HC samples. Green bullets:

NS samples.

This is a first advancement on the difficulty of

separating subsets which are much closer themselves

than with respect to a golden rule set as NS. The same

observation may be derived for the female subset,

where FPD vs FNS and FHC vs FNS are much more

divergent than between themselves (FPD vs FHC). It

may be seen that the divergences of the PD subsets

with respect to the normative set averages are much

larger than their divergences with respect to the

healthy controls. The divergences of healthy controls

with respect to the normative subsets are almost as

large than those ones from PD subsets. This

observation may indicate that the healthy controls are

farther away from normative sets than expected in

terms of articulation kinematics. On the other hand,

the normative samples are closer to their average, as

expected. These results show that PD samples are

clearly diverging from normatives, and to some

extent from healthy controls. The question now is if

this divergence is statistically significant to assume

different information contents among pathologic,

control and normative subsets. The graphical

representation of the divergence among the different

subsets may help in understanding better the

relationships involved. The divergence between each

sample in the study and the normative and control

subset averages (avMNS and avMHC) is represented

graphically in the plots shown in Figure 5.

BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing

Figure 6: Hierarchical clustering of speaker samples by JSD to their respective average normative distribution. Left: Male

subsets. Right: Female subsets.

It may be seen that the distance of the normative set

NS with respect to its centroid avNS is small, but it is

significantly larger with respect to the healthy control

centroid avHC, both for male and female samples. On

its turn, the situation of samples from healthy controls

and PD patients is the reverse, they are far from the

normative centroid avNS, but at short distance from

the healthy control average avHC. This reflects the

difficulty in separating both sets of samples as far as

vowel sustained dysarthria is concerned. This

situation is also illustrated by hierarchical clustering

in terms of each sample JSD with respect to their

respective average normative sets (avMNS and

avFNS), as reflected in figure 6.

The male set is separated into four main clusters,

one including a single sample, the second one

including six ones in which healthy controls are a

majority of 4/2 (in red), the third one grouping

normative samples and a PD sample (in green), and a

fourth one integrated by healthy controls and PD

samples in equal proportion of 4/4 (in blue). The

situation for the female subsets is a bit more complex.

There are also four main clusters, the first one

composed of three samples, two healthy controls and

a PD sample (black and cyan), the second one

integrated by the whole normative subset (one in

black and seven in purple), a third one integrated by

three healthy controls and three PD samples (in red),

and a fourth one including three healthy controls and

four PD samples. Again, the similarity between

healthy controls and PD patients is manifested in

sustained vowel dysarthria. This situation is

confirmed by t-Student, Kolmogorov-Smirnov and

Mann-Whitney tests considering equal-means null

hypothesis conditions. As it may be seen in 0, the tests

including healthy controls and PD samples vs the

normative set reject the null hypothesis, both for

males and females, pointing to strong differences

with respect to normative speakers as far as vowel

dysarthria is concerned. But the situation is

completely different when PD sets are compared with

healthy controls. Whereas for male sets t-Student and

Kolmogorov-Smirnov tests reject the null hypothesis,

Mann-Whitney fails in doing so. In the case of female

sets, all the mentioned tests fail in rejecting the null

hypothesis, pointing to more similarities than

expected between healthy controls and PD patients.

Aging voice could be behind the problem.

Table 6: Estimated p-values from inter-subset tests. t-St: t-

Student; KS: Kolmogorov-Smirnov; MW: Mann-Whitney.

The cases where the null hypothesis is not fulfilled under a

5% level are printed in bold.

Datasets

t-St

MPD vs MNS

0.000458

0.001430

0.000311

MHC vs MNS

0.000017

0.000156

0.000155

MPD vs MHC

0.0229

0.0497

0.1300

FPD vs FNS

0.000249

0.000156

0.000155

FHC vs FNS

0.000062

0.000156

0.000155

FPD vs FHC

0.366

0.188

0.195

These results show that PD datasets are clearly

separable from normative and healthy controls at

highly significant levels, both in the case of male and

female subsets. HC are also significantly different

than normative sets. But separability between PD and

age-paired HC is not granted under acceptable

standards, possibly due to the aging characteristics of

HC articulation kinematics. This is not clear in the

male set, where two tests (t-Student and Kolmogorov-

Smirnov) avail separability whereas MW does not.

But in the female case, the three tests fail in rejecting

the null hypothesis, pointing to the difficulty in

distinguishing both sets on the basis of articulation

Comparing Parkinson’s Disease Dysarthria and Aging Speech using Articulation Kinematics

kinematics, a fact which is also observable in figure

6. HC shows a closer kinematic nature with respect to

PD, which results in some confusion and separation

difficulties. It may be observed that this similarity is

of aging nature, i.e., healthiness of healthy controls

cannot be assimilated to normative articulation. Age-

paired HC show certain similarities with PD patients

due to the effects of aging in articulation, although

this assumption must be proven. A comparison of PD

datasets with respect to normative sets may be not

resolving enough, as anticipated by the kinematic

analysis of PD and presbyphonic voice. It may be

argued that articulation kinematics is not sharp

enough to establish this differentiation, but it must be

taken into account that articulation instability is quite

well modelled by AKV pdf (Gómez, P., et al., 2017).

On the other hand, studies based on phonation

features, whether linear or non-linear, should be

subject to the same three-band tests to ensure that they

are sensitive to this separability problem. These

considerations raise immediate methodological

concerns regarding tests including PD patients and

healthy controls paired in age. It is unclear if this

separability problem is due to aging voice in healthy

controls, and in that case, if distortions found in PD

samples could be due also to aging, and not only to

pathology. The conclusion is that more tests with

larger number of samples should be conducted to

confirm or reject this observation, and that sharper

methods should be designed, both for the study of

vowel and speech dysarthria, as well as for studies

involving phonation, classically based on distortion

correlates as jitter, shimmer, signal-to-noise and non-

linear features. Especial care on this respect should be

observed regarding MFCC’s (mel-frequency cepstral

coefficients), as these features are known to be

sensitive both to dysphonia and to dysarthria. In this

sense, resolutive features are to be sought and tested

using three-band benchmarks in the way shown in the

present study.

5 CONCLUSIONS

From what has been discussed, the most relevant

conclusions to be summarized are the following:

• Paired tests show that articulation kinematics is

substantially different for PD and HC with respect

to normative subjects.

• This does not seem to be the case between PD and

HC subjects, as these subsets may share some

dysarthric features which may be due to aging

more than to neuromotor degeneration.

• This differentiation problem needs to be evaluated

as well in the case of phonation features,

otherwise there will not be full guarantee in using

phonation features to assess neuromotor

degeneration.

As a final remark, it must be stressed that these

conclusions are conditioned by the low size of the

datasets used, and require further validation with

larger number of subjects to be generalized.

ACKNOWLEDGEMENTS

Funded by grants TEC2016-77791-C4-4-R

(MINECO, Spain), CENIE_TECA-PARK_55_02

INTERREG V-A Spain – Portugal (POCTEP), 16-

30805A (CZ.1.05/2.1.00/03.0072), and LOl401 from

the Czech Republic Government.

REFERENCES

Anizah, S., et al. (2018). Objective Evaluation of

Bradykinesia in Parkinson’s Disease using

Evolutionary Algorithms. Proceedings of the 11th

International Joint Conference on Biomedical

Engineering Systems and Technologies (BIOSTEC

2018): 63-69. DOI: 10.5220/0006601700630069

Buchaillard, S., Perrier, P. and Payan Y., 2009. A

biomechanical model of cardinal vowel production:

muscle activations and the impact of gravity on tongue

positioning. Journal of the Acoustical Society of

America, 126(4): 2033-2051.

Brabenec, L., et al., 2017. Speech disorders in Parkinson's

disease: early diagnostics and effects of medication and

brain stimulation, J. Neural Transm., 124(3): 303–334.

Cantiniaux, S., et al., 2010. Comparative analysis of gait

and speech in Parkinson’s disease: hypokinetic or

dysrhythmic disorders? J. Neurol. Neurosurg.

Psychiatry, 81(2): 177–84.

Dauer, W. and Przedborski, S., 2003. Parkinson's disease:

Mechanisms and models. Neuron, 39(6): 889–909.

Deller J. R., Proakis J. G. and Hansen J. H. L., 1993.

Discrete-Time Processing of Speech Signals,

NewYork, Macmillan.

Dorsey, E. R., et al., 2007, 'Projected number of people with

Parkinson disease in the most populous nations, 2005

through 2030' Neurology, 68(5): 384-386.

Dromey, C., Jang, G. O. and Hollis, K., 2013. Assessing

correlations between lingual movements and formants,

Speech Communication, 55(2): 315-328.

Endres, D. M. and Schindelin, J. E., 2003. A New Metric

for Probability Distributions. IEEE Trans. on

Information Theory; 49(7): 1858-1860.

Georgiou T. and Lindquist, A., 2003. Kullback-Leibler

Approximation of Spectral Density Functions. IEEE

Trans. on Information Theory; 49(11): 2910-2917.

BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing

Goberman, A. M., 2005. Correlation between acoustic

speech characteristics and non-speech motor

performance in Parkinson’s disease, Med. Sci. Monit.;

11(3): 109–116.

Goberman, A. M., Blomgren, M., Metzger, E, 2010.

Characteristics of speech disfluency in Parkinson

disease. J. Neurolinguistics, 23: 470-478.

Gómez, P. et al., 2009. Glottal Source biometrical signature

for voice pathology detection. Speech Communication,

51: 759-781.

Gómez, P. et al., 2017. Parkinson Disease Detection form ,

A. R.Speech Articulation Neuromechanics. Frontiers

on Neuroinformatics, doi: 10.3389/fninf.2017.00056.

Gómez, P. et al., 2018. Neuromechanical Modelling of

Articulatory Movements from Surface

Electromyography and Speech Formants. International

Journal on Neural Systems (in press), doi:

10.1142/S0129065718500399.

Gómez A., et al., 2018. Estimating Facial Neuromotor

Activity from sEMG and Accelerometry for Speech

Articulation. Proc. of the IEEE Int. Symp. on Medical

Measurements and Applications, 287-292.

Hannam, A. G., et al., 2008. A dynamic model of jaw and

hyoid biomechanics during chewing, J. Biomechanics,

41: 1069-1076.

Jankovic, J., 2008. Parkinson's disease: clinical features and

diagnosis, J. Neurol. Neurosurg. Psychiatry, 79(4):

368–376.

Jürgens, U., 2002. Neural pathways underlying vocal

control. Neurosci. and Behav. Rev. (26): 235-258.

Mardsen, C. D., 1994. Parkinson’s disease. J. Neurol.

Neurosurg. Psychiatry, 57: 672–681.

Martens, H. et al., 2015. The effect of intensive speech rate

and intonation therapy on intelligibility in Parkinson’s

disease. J. Comm. Disorders, 58: 91.105.

Mekyska, J., et al., 2015. Robust and complex approach of

patohogical speech signal analysis, Neurocomputing,

167: 94-111.

Ricciardi, L., et al., 2016. Speech and gait in Parkinson’s

disease: When rhythm matters, Park. Relat. Disord.,

32: 42–47.

Rusz, J. et al., 2013. Imprecise vowel articulation as a

potential early marker of Parkinson’s disease: effect of

speaking task, J. Acoust. Soc. Am., 134: 2171–2181.

Salicrú, M., et al., 1994. On the Applications of Divergence

Type Measures in Testing Statistical Hypotheses, J. of

Multivar. Anal. 51(2): 372-391.

Sapir, S., Ramig, L. O., Spielman, J. L. and Fox, C., 2010.

Formant Centralization Ratio: A Proposal for a New

Acoustic Measure of Dysarthric Speech, Journal of

Speech, Language and Hearing Research, 53(1): 114-

125.

Sapir, S., 2014. Multiple factors are involved in the

dysarthria associated with Parkinson's disease: a review

with implications for clinical practice and research,

Journal of Speech, Language, and Hearing Research,

57(4): 1330-1343.

Tsanas, A., et al., 2010. Novel speech signal processing

algorithms for high-accuracy classification of

Parkinson‘s disease, IEEE Trans. on Biomed. Eng. 59:

1264-1271.

Webb, A. R., 2003. Statistical pattern recognition. John

Wiley & Sons.

Comparing Parkinson’s Disease Dysarthria and Aging Speech using Articulation Kinematics