PARKINSON’S DISEASE ASSESSMENT USING SPEECH
ANOMALIES: A REVIEW
Taha Khan
Department of Computer Engineering
Dalarna University, 78188
Borlange, Sweden
[email protected]
ABSTRACT
This paper presents a detailed systematic literature review on
Parkinson’s disease severity assessment methods based on speech
impairment. Techniques that cope up the challenges of voice
signal processing in real-time have been reviewed. An analysis on
finding relevant acoustic parameters that are robustly influenced
by Parkinson’s disease symptoms and medication-related changes
was done. Time-series analysis of voice acoustics, feature
extraction and pattern recognition methods are evaluated on the
basis of portability and compatibility to mobile devices. A twotier review methodology is developed based on the applicability
of pathological voice recognition methods in real-time. The goal
is to design an algorithm based on the review synthesis that can
tackle real-time constraints in pathological voice recognition for
the assessment of Parkinson’s disease severity. Pause detection,
peak to average power rate clipping and zero thresholding rate
calculations are proposed for human voice segmentation. Previous
work depicts that these methods produce rich voice features in
real-time. We suggest that these features may be further processed
using wavelet transforms and used with a neural network for
detection and quantification of speech anomalies related to
Parkinson’s disease.
General Terms
Algorithms, Measurement, Reliability, Experimentation.
Keywords
Parkinson Disease, Hypokinetic Dysarthria, Speech Recognition.
1. INTRODUCTION
The need for real world Parkinson’s disease (PD) treatment
remains unmet for the vast majority of patients. Since PD is
progressive, there is a need of follow up treatments over time.
Medicine dosing needs to be adjusted daily with respect to
physical exercise, food intake and mood [1]. Only 3-4% of the
patients receive timely treatment [2]. Reasons for this are that
individuals may have physical limitations which make it difficult
for them to come for treatment or they may not have easy access
to treatment provided by a therapist.
High-performance personal digital assistants (PDA) with
advanced sound processing capabilities have potential to tackle
the challenge of treatment accessibility [3]. Clinician monitoring
and feedback can be preserved using the PDA since it has the
portability to adapt clinician directed treatment to home self
training[3]. Portable feedback devices for pathological individuals
have been previously investigated. Zicker et al. [3] found that an
individual with PD was able to modify her behavior when she
received feedback from a device that her speech was too soft.
Multisite treatment and the effective data to evaluate the delivery
of consistent treatment among clinicians can be acquired using a
PDA [3].
1.1 PD effects on Human Voice
PD is characterized by the loss of dopaminergic neurons in brain
[4]. This loss results in dysfunction of basal ganglia pathway
which is an essential part of the circuitry that mediates motor and
cognitive functions. As a result of dopamine loss in basal ganglia,
there can be a number of motor symptoms such as rigidity,
akinesia (loss of control over voluntary muscle movements),
bradykinesia (abnormal slowness in muscle movements), rest
tremor, postural abnormalities, and speech dysfunction [4].
Physical symptoms that can occur in the limbs can also occur in
the speech system. These symptoms are classified as hypokinetic
dysarthria (HKD) [5]. "Dysarthria" refers to a speech disorder due
to a change in muscle control. Hypokinetic means reduced
movement. Thus, hypokinetic dysarthria is reduced movement of
the muscles used for speech production.
1.2 Characteristics of Hypokinetic Dysarthria
HKD can affect respiration, phonation, resonation and articulation
in speech [4-6]. Respiration problems disturb voice loudness in
PD patients [6]. The reason is control of inhalation and exhalation
enables a person to maintain adequate loudness of speech through
a conversation. Person with PD may speak on the "bottom" of his
or her breath i.e. inhale, exhale, then speak; rather than on the
"top" i.e. inhale, speak, exhale remaining air [6]. PD individual’s
voice is an average of 2-4 dB softer than the normal voice [6].
Vocal folds vibration during phonation creates pitch of the voice.
Vocal folds vibrate quickly during high-pitched sounds and
vibrate slowly during low-pitched sounds [6]. Many individuals
with PD notice changes in pitch of their voices [7]. Monotone or
lack of vocal inflection or melody in voice is also a common
complaint[8].
Resonating system determines richness of the voice. Soft palate,
located in the back of the mouth roof closes off the nasal cavity
while speaking, except when producing nasal sounds such as
“...ing”, “m...” or "n...”. The soft palate does not move normally
in PD. A nasal quality in voice is produced as the air is leaked
into the nose due to the soft palate’s inadequate movement [8].
Articulator system comprises of the face muscles, lips, tongue,
and jaw. Imprecise articulation in PD is attributed to the reduced
movement or lack of coordination of face muscles [7]. Imagine
1.3 Research Challenges
Previous research on HKD analysis had constraints due to
laboratory controlled settings. Voice classification in real-time
environment is challenging. Speech datasets could possibly be
collected via phone calls by the PD patients so that speech
assessment can be more timely accurate and productive. But due
to the background noise or in the medium of transmission, human
voice detection becomes difficult. Distance of patient’s mouth
from the phone’s mouth-piece may also create problem in
recognizing voice amplitude [9]. Male and female voices have
different pitch properties [10]. All these issues can lead to
incorrect speech segmentation.
Enough data sets are required for training and classifying the data
according to clinical rating scale [11]. Due to the medical ethics
and patient’s privacy issue this is another challenge.
Effect of speech rate on overall intelligibility in PD is a matter of
debate. A comparison of results of previous studies on speech rate
in PD is hampered by methodological differences [12]. The
research should also contribute to prove the validity of the above
hypothesis.
The paper has been sectioned as follows; section 2 reveals the aim
of the study. Section 3 elaborates the synthesis of previous
research. An approach for the characterization of speech
impairment based on the review synthesis has been proposed in
section 4. The paper has been concluded in section 5.
2. AIM OF THE STUDY
The purpose of this study is to review methods which can
distinguish between pathological and normal human speech based
on a synthesis from a literature review. The goal of the review is
to find solutions to problems that occur in real time voice signal
processing.
[21] for speech items ranked from 0-to-4 (normal to severe
pathological state) needs to be chosen. The review methodology
has been illustrated in figure 1.
3. SYNTHESIS OF PREVIOUS WORK
In this section, the literature on acoustic analysis of HKD in PD
patient is reviewed and synthesized. Choice of publications for
review is based on the validation of methodology, credibility of
experimental techniques for real-time voice analysis and
portability of algorithm for the mobile companion. Synthesis is
presented in form of a time-line i.e. starting from the investigation
of essential acoustic parameters for HKD recognition to the
methods implemented for classification between HKD from the
normal voice.
Keywords such as “Parkinson’s disease”, “Hypokinetic
Dysarthria”, “Speech Recognition in real time” were used in the
search engines. Histogram (in figure 2) displays the search
engines used, number of total relevant literature found and the
filtered publications based on review methodology. The most
useful search hits for the synthesis have been found using the
IEEE Explorer search engine. Other search engines used were
Google Scholar, ELIN, EBRARY, and LIBRIS. Below is the
synthesis of relevant literature evaluated.
Database Search
250
Num ber of Hits
that your face was very cold and it was difficult to move your
facial muscles, your speech becomes slurred and unclear [6].
200
150
Relevant Hits
100
Useful Hits
50
0
Google
Scholar
ELIN
Ebrary
LIBRIS
IEEE
Explorer
Search Engines
Figure 2. Search for the Relevant Literature.
3.1 Acoustic Measures for HKD recognition
Investigation of relation between the pathological acoustic
parameters and Parkinson’s disease has been of primary concern
for researchers. Kristin M. Rosen, et al. [5] identified acoustic
signatures that were robust to phonetic variation in PD patients
speech conversations. 20 healthy control (HC) and 20 PD patients
had 2 minutes of conversational speech. Group differences were
detected in T-test in 8 of the 9 measures. Most specific (95%) and
accurate (95%) differentiators of HKD and HC speech were found
to be ‘Pause Time’ and ‘Spectral Range’. It was concluded that
HKD can be distinguished from HC voice on the basis of
amplitude variation and spectral range.
Figure 1. Review Methodology.
Acoustic features that are robust to medication changes in patients
need to be determined. A choice of an efficient time-series
analysis tool for voice signals needs to be made for better
evaluation of segmented speech. A classification tool to classify
pathological speech according to clinical speech assessments
using the UPDRS (Unified Parkinson’s Disease Rating Scale)
An approach was adopted by Paul McRae and K. Tjaden to
characterize the relationship between vocal tract acoustic output
and perceptual impressions of PD patient’s speech [13]. In this
method speakers read the passage with habitual slow and fast
rates. Voice of vowels ‘i’, ‘u’ and ‘a’ and fricatives ‘f’ and ‘s’
were of interest. Acoustic measures included articulatory rate,
segment durations, vowel formant frequencies, and first moment
coefficients. Results indicated high variations in temporal acoustic
measures for the PD group when compared to HC group.
Kris Tjaden [14] used vowel co-articulation and vowel syllables
to study acoustic variability between 9 HKD males with PD and a
group of 10 HC males [14]. Graded speaking rate was used to
investigate effect of rate variation on co-articulation. Ratio of F2
formant onset frequency and F2 target frequency was used to infer
co-articulation. Fricative F2 was obtained for speech stimuli and
compared to F2 onset measures. The result (shown in figure 3a
and 3b) was, ratios tended to be smaller for PD speakers than for
HC [14].
frequencies [7]. To improve differentiation of HKD from normal
speech, the acoustic metric must be minimally affected by
speaker-related variability and maximally affected by the
articulatory impairment, as reflected by vowel formant
centralization. The metric used was called the formant
centralization ratio (FCR). FCR effectively and robustly
differentiated the groups with dysarthria from the HC group in
[7]. Also, the FCR is insensitive to gender effects. FCR
correlated highly when the correlated variable was the change
induced by treatment. These findings suggest that the FCR is a
valid and highly sensitive metric of vowel articulation for normal
as well as for abnormal speech.
3.3 Automatic HKD Recognition
a. Two repetitions of word
“dice” are shown. Dashed
trajectory corresponds to a
repetition for PD speaker.
Solid trajectory corresponds
to repetition for healthy
speaker of similar age.
b. Two repetitions of word
“dice” illustrate varying coarticulation. Solid repetition
was produced at relatively
faster speaking rate and has
shorter vowel duration; dashed
repetition was produced at
slower rate and has longer
vowel duration.
Above studies narrowed down the research to formant frequency
of consonants and vowels as it contains most relevant acoustics
for HKD analysis. Emily Budkowski [8] calculated PD severity
ratings using fundamental frequency of Voice Onset Time (VOT
is the interval between the initial articulatory release of a stop
consonant and onset of voicing for the subsequent vowel).
Participants’ speech ranged from mild to severe. He found out
that Levodopa appeared to have greater effect on VOT within PD
group and it is a measure of medication related rate change.
Results indicated that all participants exhibited HKD in OFF
medication state.
Figure 3. Formant Analysis [14].
Jennifer Kleinow and Anne Smith [15] revealed that for all
speaker groups (HC and HKD), STI (Spatiotemporal index, a
measure of spatial and temporal variability) values from the loud
speech were closest to those from habitual speech. The reason is
normalized movement pattern for loud speech resembled that of
habitual speech. Amy T. Neel [16] and Kris Tjaden [17] depicted
that loud speech resulted in intelligibility improvement in HKD
acoustics. Above studies narrowed down the focus of research to
vowel formant time-frequency analysis to classify between
pathological and normal voice.
a. Control group
b. Patients group
Figure 4. Sound Power Analysis [10].
Izworski et al. [10] worked with continuous sound analysis of PD
patients for automatic HKD recognition. Sound emitted for a long
period of time allowed for sound power analysis. Power value of
each frame is calculated by summing up the values of the signal
energy within the respective x (a) and x (a+m) limits of the t frame
given in equation 1.1 where m is the frame length.
a +m
3.2 Vowel Formant Frequency
The most relevant acoustic parameters for production of vowels
are the frequencies of first two formants, F1 and F2 [7]. These
formant frequencies change as a function of the movements of
articulators. In general, the frequency of F2 increases, and that of
F1 decreases, as the tongue moves forward (e.g., to form the
vowel ‘i’), and the frequency of F2 decreases as the tongue moves
backward (e.g., to form the vowels ‘u’ and ‘ao’). Also, the
frequency of F1 decreases when the tongue is elevated (e.g., to
form the vowels ‘i’ and ‘u’) and increases when the tongue is
lowered, alone or with a downward movement of the jaw (e.g., to
form the vowel ‘a’). Furthermore, the frequencies of both F1 and
F2 decrease when the lips are rounded (e.g., to form the vowel
‘u’) and increase when the lips are retracted or become unrounded
(e.g., to form the vowels ‘i’ and ‘a’).
HKD results in vowel formant centralization, that is, formants that
have high frequencies tend to have lower frequencies, and
formants that have low frequencies tend to have higher
P(t ) = ∑ x(k )
2
(1.1)
k =a
Polynomial p(x) of degree 4 that fits the data, p(x (t)) to P (t) in
least squares sense is calculated to represent average values of P
(t) is given in equation 1.2.
p( x) = p1 x n + p2 x n −1 + ... + p n x + p n+1
(1.2)
Sum of differences between vector P (t) and p(x) can be used as a
voice stability parameter which is given as.
a+m
Stab = ∑ P(k ) − p(k )
(1.3)
k =1
Values obtained were compared with the values obtained in the
control group as shown in figure 4a and 4b. Values of Stab
parameter were much greater for patients (i.e. Stab>0.6) than for
persons from the control group. With many patients, distinct and
varying breaks in phonation were observed as shown in figure 4b.
With healthy persons, gradual quietening occurred, whereas PD
patients ended the emission abruptly. This audio signal is
transformed into Fourier series and it is found out that changes are
observable especially in the vowels articulation. Results constitute
the beginning of tests concentrated on automatic voice
classification [10].
Artificial Neural Networks (ANN) can be used as a statistical tool
to formulate the boundaries between speech abnormalities using
the voice features. Mehmet F. et al. [18] used ANN for feature
selection of pathological voice datasets. A fuzzy logic based ANN
(called ANFC: Adaptive Neuro-Fuzzy Classifier) gave best
recognition results. Datasets for this study consisted of 195
sustained vowel phonation from 31 people, of which 23 were
diagnosed with PD. Pathological voice features such as
fundamental frequency of voice, shimmer and jitter were used for
classification. ANFC produced 94.35% accuracy for classification
of PD group dataset.
Using Discrete Wavelets Transforms (DWT) a clear difference
was noticed between wavelet evolutions of HKD and normal
voice [19]. A visual pattern is shown in figure 6a and 6b.
ANN classification with five DWT coefficients produced 90%
classification rate for pathological voice and 100% classification
rate for normal voice [19]. Experiment proved that the formant
analysis methods using Fourier Transform is not a suitable tool
for speech analysis [11]. Since speech is a highly non-stationary
signal, wavelet transform proved to be a better tool for analysis of
non stationary signals [19] as it is useful in localizing a symptom
both in time and frequency scales.
3.4 Automatic HKD recognition using hybrid
approach: Wavelet Analysis and ANN
L. Salhi et al [19] used Linear Predictive coding method on PD
patients voice on formants F1, F2 and F3. Comparing with the
normal voice, pathological voice showed high variations.
a. Normal Voice
b. Pathological Voice
Figure 6. Wavelet Analysis [11].
Above synthesis reveals that the assessment of PD is possible
through speech datasets via reduced voice loudness, mono pitch,
short rushes, prolonged syllables, long pauses, and reduced
phonation time. The synthesis depicts that HKD effected PD
patients can be effectively classified on the basis of UPDRS
ratings using classification tool like ANN.
4. PROPOSED ALGORITHM
The biggest drawback found in the previous work was the
controlled setting of experiments in laboratory environment [9,
20]. Since this study is aimed for real time processing of speech
signals in a mobile companion, the biggest challenge is voice
segmentation and noise removal in the speech signals achieved
during phone calls in home environment. An algorithm design is
proposed to tackle this challenge. First pre-processing of voice
signals needs to be done succeeded with voice segmentation. The
segmented voice may be decomposed using wavelet and later
trained in the neural network for classification. Below are the
details of each step.
Figure 5. A Hybrid approach for HKD recognition [11].
Although these methods distinguished HKD with the normal
voice but the problem was, these methods do not produce
quantification values to make the decision. They introduced a
hybrid approach using wavelet analysis and ANN. Normalized
energies and entropies of wavelet coefficients were used to
formulate feature vector of speech sample. The feature vector was
used to identify impaired speech. A 3-layered feed forward ANN
with Back Propagation (BP) algorithm was used for classification.
The flowchart of this algorithm is shown in fig. 5.
4.1 Voice Pre-processing
Most intelligibility relevant speech segments (consonants and
vowels) contains minority of speech energy [9]. Intelligibility can
be improved by amplifying if total energy remains constant.
Amplitude compression is a dynamic function which provides
larger amplification when input has low power over certain time.
The effect is that, power of sound varies less in time and hence it
is less noisy [9]. Speech enhancements can be done using High
Pass filter and clipping [9]. Also the speech signal does not have
unique maximum amplitude. Some waveforms are asymmetric
than others, that is, they reach higher peaks in positive amplitude
than negative or vice versa. More asymmetric waveforms have
larger peak power. Human speech has “Peak to Average Peak
Power” (PAPR) [9]. In order to remove noise, peaks over or less
than PAPR threshold range can be clipped off in the speech
signals. The speech signal can be further “windowed” using a
Hamming Window [9].
Once the signal is pre-processed, voice can be segmented based
on vocal sounds. Syllable segmentation of speech is important for
correct voice classification because it reduces search space
effectively.
An automatic processor should succeed at identifying silence,
vocalic, fricative and nasal sounds and stop gaps to provide gains
in speech intelligibility [19]. In case of HKD, segmentation is
difficult because syllable units are spread roughly via intensity
changes but exact boundary positions are elusive in successive
vowels.
4.2 Decomposition with Wavelet Transforms
Once the voice signal is segmented, speech can be decomposed in
transient and tonal components using Discrete Wavelet transforms
(DWT) [11].
In disordered speech, the non-stationary behavior of the pitch can
be analyzed using wavelets. The criteria for selecting a proper
mother wavelet is to have a wavelet function with enough number
of vanishing moments in order to represent the salient features of
the disturbance. At the same time, this wavelet should provide
sharp cut-off frequencies. Selected wavelet should be ortho
normal. Daubechies 40 (Db40) shows sharper cut off frequency
compared with the others. Using Db40, leakage energy between
different resolution levels can be reduced [11].
Wavelet
transforms outputs energy coefficients that represent the signal in
time and frequency. Since wavelet energy coefficients are
sensitive to transients, areas of changes can be amplified and
detected using Laplace derivative. This can make the speech
transients very rich [9].
Amplitude Compression
Acquire Speech
High-Pass Filter
Figure 7. Five possible pause points in an utterance [20].
The first step in voice segmentation is “Pause Detection”. A pause
exists at the end of several continuous speech words [20]. Each
character is a syllable and each pause indicates syllable boundary,
therefore pause detection plays a very important role in syllable
segmentation. Short term energy in voice signals can be used to
detect pauses. A threshold of short term energy is set first to
detect pause. If the energy is less than certain threshold, it is
marked as a pause and a syllable boundary [20]. After pause
detection, speech signal may be divided into pieces (fig. 7).
Voice Preprocessing
PAPR Clipping
Windowing
Voice Segmentation
Pause Detection
Wavelet Transform
ZCR threshold of Formants F1
and F2
ANN
UPDRS 5.4(Unintelligible)
Table 1. Voice Segmentation based on ZCR analysis [20].
With ZCR Rules
Without ZCR Rules
Syllables
856
856
Correct
819
803
Deletion
37
53
Insertion
64
79
Correct Rate
95.7%
93.8%
Error Rate
11.8%
15.4%
For continuous speech segmentation, zero crossing rates (ZCR)
provide spectral information at low computational cost. ZCR is
the rate when a waveform crosses time axis or changes its
algebraic sign. Consonants have high ZCR whereas vowels have
low ZCR [20]. Therefore an onset of high ZCR means beginning
of consonant and it should be a starting boundary. An offset of
high ZCR means an endpoint of consonant and it should be an
ending boundary [20].
Table 1 depicts the comparison between voice segmentation with
ZCR rules and without ZCR rules. Segmentation accuracy is
drastically improved when ZCR rules are applied to the voice
segmentation algorithm as shown in table 1.
Decision
UPDRS 5.0 (Normal)
UPDRS 5.3 (Severe)
UPDRS 5.1 (Mild)
UPDRS 5.2 (Moderate)
Figure 8. Flowchart of Proposed Algorithm.
4.3 Classification Using ANN
Resulting DWT coefficients can be used as feature vector for
training artificial neural networks using Back Propagation
algorithm. In order to obtain optimal results, nature of the input
coefficients vector is to be varied for training in ANN.
As suggested by L. Salhi et al.[19], three different input vectors
can be used for ANN. They are (1) DWT coefficients (2) DWT
energy coefficients and (3) DWT entropy coefficients. The
resulted pathological voice can be further classified according to
UPDRS ratings [21] Based on the classification results, the level
adjustments of medication can be made and other likely
treatments can be opted. The flowchart of the proposed algorithm
is shown in figure 8.
5. CONCLUSION
This review evaluated HKD recognition algorithms to develop a
tool for PD severity assessment using speech datasets of PD
patients. Previous research had constraints due to the controlled
settings of the laboratory environment. Possible solutions for
voice analysis in real-time have been investigated. Voice acoustic
parameters that were robust to medication changes are evaluated.
Since voice signals are highly non-stationary, focus was laid on
quantization of speech signals. During the study it was found that
time-frequency methods fail to quantify the voice signals over the
time-series based on the voice frequency. Wavelet transformation
is found promising for voice analysis because it quantizes the
non-stationary voice signals over the time-series using scale and
translation parameters. In this way voice intelligibility in the
waveforms can be analyzed in each time frame. During the
review, artificial neural network was found to be a promising tool
for voice classification.
This review is a basis for future work to conceive a HKD
recognition tool in a PDA for self-treatment of PD patients. Also
clinicians will be aided from feedback generated by this tool in
follow up of patients who suffer from HKD.
6. ACKNOWLEGDEMENT
This research is a module of the project “E-Motions”. This project
has been running in Dalarna University in collaboration with
Abbott Laboratories and Nordforce Technology and is funded by
a grant from Swedish Knowledge Foundation.
7. REFERENCES
[1] Stewart, A. F. and Weiner, W.J. 2008.Parkinson's disease:
diagnosis and clinical management. Demos Medical
Publishing, New York.
[2] Hartelius, L. and Svensson, P. 1994. Speech and swallowing
symptoms associated with Parkinson's disease and multiple
sclerosis. Folia Phoniatr Logop. 46, 1(1994), 9-17.
[3] Zicker, J., Tompkins, W., Rubow, R., and Abbs, J. A
Portable Microprocessor-Based Biofeedback Training
Device. IEEE Trans Biomed Eng. 27 (Sep. 1980), 509-515.
[4] Joan, K.Y., Tara,M., Whitehill, L., and So, S.Y.S. Intonation
Contrast in Cantonese Speakers with Hypokinetic Dysarthria
Associated with Parkinson’s disease. J Speech Lang Hear
Res. 53(2010), 836-849.
[5] Rosen, K. M., Kent, R. D., Delaney, A. L., and Duffy, J. R.
Parametric Quantitative Acoustic Analysis of Conversation
Produced by Speakers with Dysarthria and Healthy Speakers.
J Speech Lang Hear Res. 49 (2006), 395-411.
[6] Stefanie, C. and Schwantz, J.1998. Parkinson's disease:
Speaking Out. The National Parkinson Foundation, Inc,
Denver, Colorado.
[7] Sapir, S., Ramig, L.O., and Spielman, J.L. Formant
Centralization Ratio: A Proposal for a New Acoustic
Measure of Dysarthric Speech. J Speech Lang Hear Res. 53
(2010), 114-125.
[8]
Budkowski, E. 2007. Voice Onset Time in Parkinson
Disease. Master’s thesis. Bowling Green State University.
[9] Dominguez, A.C.2009. Pre-Processing of Speech Signals for
Noisy and Band-Limited Channels. Master’s thesis. KTH,
Stockholm.
[10] Izworski, A., Augustyniak, P., and Orzechowski, T. 2006.
Processing and analysis of voice anomalies in course of
Parkinson's diseases. In Proceedings of Eighth IASTED
(Honolulu, Hawaii, USA, 14-16 August, 2006), 354-357.
[11] Salhi, L., Mourad, T., and Cherif, A. Voice Disorders
Identification Using Multilayer Neural Network. IAJIT, 7, 2
(2010), 177-185.
[12] De Letter, M., Santens, P., de Bodt, M., Boon, P. and
Borsel, J. V. Levodopa induced alterations in speech rate in
advanced Parkinson’s disease. Acta Neurol Belg, 106 (2006)
19-22.
[13] Salhi, L., Talbi, M., and Cherif, A. Voice Disorders
Identification Using Hybrid Approach: Wavelet Analysis and
Multilayer Neural Networks. WASET. 45 (2008) 330-339.
[14] McRae, P., and Tjaden, K. Acoustic and Perceptual
Consequences of Articulatory Rate Change in Parkinson
Disease. J Speech Lang Hear Res. 45(2002), 35-50.
[15] Tjaden, K. An Acoustic Study of Co articulation in
Dysarthric Speakers with Parkinson Disease. J Speech Lang
Hear Res. 43 (2000), 1466-1480.
[16] Kleinow, J., Smith, A., and Ramig, L.O. Speech Motor
Stability in IPD: Effects of Rate and Loudness
Manipulations. J Speech Lang Hear Res. 44 (2001) 10411051.
[17] Neel, A.T. Effects of Loud and Amplified Speech on
Sentence and Word Intelligibility in Parkinson Disease. J
Speech Lang Hear Res. 52 (2009), 1021-1033.
[18] Tjaden, K., and Wilding, G.E. Rate and Loudness
Manipulations in Dysarthria: Acoustic and Perceptual
Findings. J Speech Lang Hear Res. 47 (2004), 766-783.
[19] Caglar, M.F., Cetisli, B., and Toprak, I.B. Automatic
Recognition of Parkinson’s disease from Sustained
Phonation Tests Using ANN and Adaptive Neuro-Fuzzy
Classifier. J Speech Lang Hear Res. 1, 2 (2010), 59-64.
[20] Zhao, X., and O’Shaughnessy, D. 2008. A new hybrid
approach for Automatic Speech Signal Segmentation using
silence signal detection, energy convex hull and spectral
variation. In IEEE CCECE, (Canada, 2008) 145-148.
[21] Fahn S et al. 1987. UPDRS. Unified Parkinson Disease
Rating Scale. In Fahn S, Marsden CD, Calne DB, Goldstein
M, eds. Recent developments in Parkinson’s disease.
(Florham Park, New Jersey: Macmillan Healthcare
Information, 1987), 153-163.