2008 International Conference on Signals, Circuits and Systems

Voice Disorders Classification Using

Multilayer Neural Network
Lotfi Salhi Talbi Mourad Adnéne Cherif
Signal Processing Laboratory - Sciences Faculty of Tunis - Tunis ElManar – Tunisia
lotfi.salhi@laposte.net mouradtalbi1969@yahoo.com Adnane.cher@fst.rnu.tn

we attend confusion between the modifications of

Abstract—In this paper we present a new method
for voice disorders classification based on multilayer
neural network (MNN). The processing algorithm is
based on a hybrid technique which uses the wavelets
energy coefficients as input of the MNN. The
training step uses a speech database of several
pathological and normal voices collected from the
national hospital “Rabta - Tunis” and was
conducted in a supervised mode for discrimination
of normal and pathology voices and in a second step
classification between neural and vocal pathologies
(Parkinson, Alzheimer, laryngeal, dyslexia…).
Several simulation results will be presented in
function of the disease and will be compared with
the clinical diagnosis in order to have an objective
evaluation of the developed tool.
I. INTRODUCTION
Pathological voice recognition has been received
a great attention from researchers in the last decade.
Speech processing has proved to be an excellent
tool for voice disorder detection. Among the most
interesting recent works are those concerned with
Parkinson’s Disease (PD), multiple sclerosis (MS)
and other diseases which belong to a class of neuro-
degenerative diseases that affect patients speech,
motor, and cognitive capabilities [1,2].
the motivity orofacial and the speech quality that
remains the clinic objective [7].
Features assessment of a voice disorder is that the
disorder carries on a patient's capacity to
communicate are a crucial step to conceive a
program of its management. A process of the
prosperous assessment allows the pathologist of the
speech to diagnose the voice disorder, determine
the relative efficiency of several treatment
approaches and formulate a prognosis [8].
Physicians often use invasive techniques like
endoscopy to diagnose the symptoms of vocal fold
disorders. However, it is possible to identify
disorders using certain features of speech signals
[3,4].
Different classic techniques are used to extract the
vocal parameters and so to make the classification
of the pathological voices such as pitch and formant
detection: PDA (cepstrum, FFT, spectrogram …).
II. DISORDER IDENTIFICATION BY PITCH AND
FORMANTS ANALYSIS
A. Speech Processing Algorithm
Speech Acquirement

The speech production is a complex motor act that

implies a big number of muscles, of physiological Pre-emphasis

variables and a neurological control implying

different cortical and under cortical regions. Segmentation by
Hamming Window
We distinguishe three systems contributing to the
production of the speech: the respiratory system, Pitch Detection Formants extraction
the laryngeal system and the supra-laryngeal system by Cepstrum by LPC

(the articulators) [9,10].

The nervous system also controls the prosody. This F0 F1, F2, F3
one schematically covers the variations of height
(intonation, melody), the variations of intensity
(accentuation) and the temporal progress (pauses, Comparison with
Normal Values
debit, rhythm).
The analysis of the voice disorder stays essentially
clinic [3]. The instrumental measures are spilled Decision

little in practice clinic. The most used are the

acoustic and aerodynamic measures [4]. The speech
analysis is complex and has been disregarded for a Pathological Healthy

long time. A difficulty result is to analyze in the

literature the different treatment effect (medical or Neural Organic

surgical). Indeed, a many studies don't return

specific analysis of the speech. On the other hand, Figure 1: Speech processing algorithm

978-1-4244-2628-7/08/$25.00 ©2008 IEEE -1-

2008 International Conference on Signals, Circuits and Systems

B. Pitch And Formants Analysis Results
Figure 2 illustrates the pitch variation by
application of the cepstrum method analysis of
normal and pathological female sounds (32 years).
Figure 2-a: Pitch evolution for a normal female voice
take the decision. For it, we try to improve this idea
by new methods that use wavelet transformed and
neural networks.
III. NEW APPROACH FOR VOICE PATHOLOGY
CLASSIFICATION
This work presents a development of the basic idea
presented in [5]. This paper propose a technique
that uses wavelet analysis to extract a feature vector
from speech samples, which is used as input to a
Multilayer Neural Network classifier. Wavelet
analysis provides a two-dimensional pattern of
wavelet coefficients. The energy content of
Wavelet coefficients at various level of scaling is
used to formulate a feature vector of speech sample.
Attempt is made to use this feature vector as a
diagnostic tool to identify pathological disorders in
the voice. A three layer feed forward network with
sigmoid activation is used for classification.
Generalized Back Propagation Algorithm (BPA) is
used for training of the network.

A. Algorithm Of The Hybrid Method

Speech Acquirement

Figure 2-b: Pitch evolution for a pathological female voice Pre-emphasis

The linear predictive coding (LPC) method Segmentation by

applied on the same voices allows evolution to Hamming Window
extract the formants. By comparison to the normal
Wavelet Transform
values (figure 3), the high variations of the
formants F1, F2 and F3 of the pathological male
Energy
sound confirm the conclusions. [12,13] Coefficients

Multilayer Neural
Network (MNN)

Decision

Pathological Healthy

Figure 4: Hybrid method algorithm

B. Wavelet transforms analysis

Figure 3-a: Formants evolution for a normal male voice
The wavelet transform can be viewed as
transforming the signal from the time domain to the
wavelet domain. This new domain contains more
complicated basis functions called wavelets, mother
wavelets or analysing wavelets.
A wavelet prototype function at a scale s and a
spatial displacement u is defined as: [6]
1 t  u ( u  IR, s  IR* ) (1)
\ (t )
u,s \( ) 
s s

C. Continuous Wavelet Transforms

Figure 3-b: Formants evolution for a pathological male voice
Although these methods can help us to distinguish a
pathological voice but they remain subjective
methods that don't give any quantification values to
-2-
The CWT is an excellent tool for mapping the
changing properties of non-stationary signals. The
CWT is also an ideal tool for determining whether
2008 International Conference on Signals, Circuits and Systems
or not a signal is stationary in a global sense. When
a signal is judged non-stationary, the CWT can be
used to identify stationary sections of the data
stream.
Specifically, a Wavelet Transform function
f(t) L2(R) (defines space of square integrable
functions) can be represented as:
f
W ( f )(u, s ) ³f
f (t )\ u*, s (t )dt
f 1 t u
³
f
f (t ) \ * (
s s
)dt
The factor of scale includes an aspect transfer at a
time in the time brought by the term u, but also an
aspect dilation at a time in time and in amplitude
brought by the terms s and s et
D. Discrete Wavelet Transform
The Discrete Wavelet Transform (DWT), which is
based on sub-band coding is found to yield a fast
computation of Wavelet Transform. It is easy to
implement and reduces the computation time and
resources required.
The Discrete Wavelet Transform (DWT) involves
choosing scales and positions based on powers of
two, so called dyadic scales and positions. The
mother wavelet is rescaled or dilated by powers of
two and translated by integers.
The Daubechies wavelet db40 is selected as a good
choice because of its high performance in an
informal listening test [15]. For fast wavelets
transformed (DWT), the functions are defined by a
game of indications that one designates under the
appellation " coefficients of the filters in
wavelets"[6].
Figure 5 shows the variation of the analysis
function of the Daubechies wavelet (db40) and its
corresponding spectrum.
Psi function of Daubechies Wavelet
0.8
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
0 20 40 60 80
time(ms)
Figure 5: Wavelet of Daubechies to 40 hopeless moments and its
spectrum
E. DWT and Filter Banks
Starting with a discrete input signal vector x[n], the
first stage of the fast wavelet transform (FWT)
algorithm decomposes the signal into two sets of
coefficients. These are the approximation
coefficients cA1 (low frequency information) and
the detail coefficients cD1 (high frequency
information). The DWT is computed by successive
lowpass and highpass filtering of the discrete time-
domain signal as shown in figure 6.
(2) Figure 6: Three-level wavelet decomposition tree
In the figure, the signal is denoted by the sequence
x[n], where n is an integer. The low pass filter is
denoted by G0 while the high pass filter is denoted
by H0. At each level, the high pass filter produces
detail information, d[n], while the low pass filter
associated with scaling function produces coarse
s. approximations, a[n].
F. Neural Network
Neural were chosen as a method of pattern
matching for many main reasons. First the MatLab
software has a fantastic implementation of several
different types of neural networks in its Neural
Network Toolbox. The big advantage of the neural
networks resides in their automatic training
capacity, what permits to solve some problems
without requiring to the complex rule writing, while
being tolerant to the errors [11].
A neuron (Figure 7) is an information processing
unit, which is an essential part of a neural network.
A neuron consists of three main elements: synapses
(links), a linear combiner, and an activation
function. Each synapse (link) contains a weight
factor. Input p(i), which is connected to neurone k,
is multiplied by synaptic weight w(k, i).
Hence, the output of a neuron depends on its inputs
and its activation function. There are different types
of activation functions that can be used in Matlab.
The most commonly used activation functions are
Specter of Daubechies Wavelet
hard limit, linear, or sigmoid functions. Naturally,
1.4
one can also construct hes own activation function.
1.2
1
0.8
0.6
0.4
0.2
0
0 5000 10000
frequency (Hz)
Figure 7: Model of an artificial neuron
Each layer has a weight matrix W, a bias vector b,
and an output vector a (Figure 8). The number of
neurons usually varies between each layer. In
Figure 3, the number of inputs is R, and the number
of neurons in the first layer is S1, while in the
second layer it is S2, also the same for other layers.
The layers, which are situated between the inputs
and the output layer, are called hidden layers. Thus,
Figure 8 shows two hidden layers [11].
-3-
2008 International Conference on Signals, Circuits and Systems

illness. The simulation result of different absolute

coefficients is given in figure 10 and 11.
We notice a clean difference between wavelet
coefficient evolutions of the two different signals.
This analysis method can also provide a visual
pattern, which can be of considerable help in
diagnostics.
Analyzed signal of normal voice.
0.5

Figure 8: Multi-layer neural network structure -0.5

0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000

A neural network is trained by giving a target Discrete Transform, absolute coefficients.

output to a certain input group, in which case the

level
term supervised learning is used. Alternatively, a
1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
network can be trained through self-guidance, Continuous Transform, absolute coefficients.

which means that the network parameters adapt 31

29
27
25
23
21

Scale
19
17
according to the input. In both cases, the free 15
13
11
9
7
5
3
parameters in the network, weights and biases, 1
1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
time (or space) b
adapt according to the measured data. The training
Figure 10: Wavelet Analysis of Normal Voice
can be gradual (incremental training), which means
Analyzed signal of pathological voice.
that the weights and biases are adapted every time 1

that a new training example is fed to the network, 0

or it can be done in batches (batch training), in -1

0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
which case the parameters are not adapted until all Discrete Transform, absolute coefficients.

the examples have been fed.

level

The backpropagation algorithm is often used in the

training of multi-layer neural networks. 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Continuous Transform, absolute coefficients.

If sigmoid activation functions are used in the 31

29
27
25
23
21
Scale

19
output layer, the outputs of the network are limited 17
15
13
11
9
7
5
3
to a small range. Also, if a linear activation function 1
1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
time (or space) b
is used in the output layer, the network outputs can
Figure 11: Wavelet Analysis of Pathological Voice
have any real number values.
B. Neural Network Design
IV. SIMULATION RESULTS
In our survey we use a multilayer neural network
The Matlab7.0 platform is used for implementation
(MNN) with only one layer hidden between the
of the neural network formed of three layers, one of
input layer and the output layer. Every neuron of
input, one of output and a hidden layer (Figure 11).
the hidden layer is connected to the neurons of the
The input layer is formed of the same neurons
input layer and those of the output layer and there
number that corresponds to the components of the
is not a connection between the cells of a same
input vector. The input is the feature vector
layer. The activation functions used in this type of
obtained from Wavelet decomposition.
network are the doorstep or sigmoid functions.
The hidden layer contains fifteen neurons and the
This network follows a supervised training
output layer contains only one neuron to give the
according to the rule of errors correction. The
decision pathological or normal (Figure 9).
training type used for this network is the
supervised fashion. To every well stocked input
an answer corresponds waited at the output. So
the network is going to alter until it finds the good
output.
C. Feature Extraction
A Filter Bank is used to extract the wavelet
coefficients.
The energy of every level is normalized against
Figure 9: Voices classification model total energy content in the signal.
E (3)
A. Wavelet Coefficients E (i) = i N
ET
We apply discreet and continuous wavelet
transform coefficients on the same word Where i = 1, 2 …
pronounced by two speakers that have the same ET : Total Energy across all the levels.
sex and same age. One of these speakers is E T = ¦ Ei (4)
healthy whereas the other sulfur of Alzheimer's i

Ei : Energy at each level.

-4-
2008 International Conference on Signals, Circuits and Systems

Table I: Neural Network Results With Three Coefficients

V. NEURAL NETWORKS RESULTS
We use 60 words on the total, pronounced by Pronounced Normal Pathological
different speakers of which 30 are normal and the word
other present pathologies of vocal or neurological training
20 20
origin. For the training, we use 40 words (20 number
normals and 20 pathological). Test number 10 10
After training the network will be tested with 20 Correcte
9 8
words different from those used for the training classification
(10 normals and 10 pathological). Rate of
90 % 80%
In order to obtain optimal result, we vary at every classification
time the number of the energy coefficients to the
input of the neural network. This procedure B. Five Energy Coefficients
requires a variation of the choice of the waveket We use a wavelet filter bank to extract a five
filter bank. wavelet coefficient then we calculate their
corresponding energy
A. Three Energy Coefficeients The results of neural network training and testing
We use a wavelet filter bank to extract a three are regrouped in tableII.
wavelet coefficient then we calculate their
corresponding energy Table II: Neural Network Results With Five Coefficients

Tree Decomposition Pronounced Normal Pathological

Word
(0,0)
Training
20 20
Number
Test Number 10 10
Correcte
(1,0) (1,1) 10 9
Classification
Rate of
100 % 90 %
Classification

(2,0) (2,1)
C. Seven Energy Coefficeients
We use a wavelet filter bank to extract a seven
wavelet coefficient then we calculate their
E1 E2 E3 corresponding energy
The results of neural network training and testing
Figure 12: Three energy coefficients extraction
are regrouped in table III.
The obtained training curve is given in figure 13 Table III: Neural Network Results With Seven Coefficients
Performance is 9.96138e-006, Goal is 1e-005
1
10 Pronounced Normal Pathological
0
10
Word
Training
-1
10 20 20
Number
Training-Blue Goal-Black

-2
10
Test Number 10 10
-3
10
Correcte
10 10
-4
10 Classification
-5
10
Rate of
-6 100 % 100 %
10
0 20 40 60 80 100 120 140 160 180 200 Classification
202 Epochs

Figure 13: Training curve with three coefficients Then we can resum these different results in the
following diagrams (figure 14).
The results of neural network training and testing
are regrouped in table I.

-5-
2008 International Conference on Signals, Circuits and Systems
Pathological Voices Classification
100
80
60
40
20
0
1 2 3 4
E n e r g y C o e f f i c i e n t N u mb e r
Normal Voices Classification
100
80
Classification Rate
60
40
20
0 decomposition: Wavelet representation” , IEEE
1 2 3 4
Energy Coefficient Number
Figure 14: Voice classification by MNN results
VI. CONCLUSION
The goal of this work is to conceive a tool of help
to the clinicians in the Tunisian hospitals. This tool
allow to follow-up of patients who suffer from
illness of vocal and neurological origin.
We presented in this paper a material and software
interface of numeric treatment of the patient’s vocal
signal based on neural networks.
Result of the multilayer neural network (MNN)
classifier gives the correct classification. The
classification rate is between 90% and 100%. We
have demonstrated in this study, a feature vector
based on wavelet coefficients is useful for
classification of normal and pathological speech
data. At a preliminary level, the speech data is
classified into two classes normal or pathological.
The multilayer neural network (MNN) with back
propagation algorithm (BPA) used as a classifier
has been proved to be more efficient and more
precise than the time-frequency analysis method.
The MNN classifier represents a low cost, accurate,
and automatic tool for pathological voice
classification using wavelet coefficients normalized
energy. It is presented in this paper as diagnostic
tools to aid the physician and clinician in the
analysis of speech disease.
Therefore, future work will be focused on the
specific recognition of illness type that causes the
speech pathology.
Finally, This work has to be validated on a larger
speech pathology database to increase the result
reliability.
VII. REFERENCES
[1] V. Parsa and D. G. Jamieson, “Interactions
between speech coders and disordered speech,”
Speech Communication, vol. 40, no. 7, pp. 365–
385, 2003.
[2] S. B. Davis, “Acoustic characteristics of normal
and pathological voices,” Speech and Language:
Advances inBasic Research and Practice, vol. 1, pp.
271–335, 1979.
5 6 7 [3] F. Plant, H Kessler, B Cheetham, J Earis,
“Speech Monitoring of Infective Laryngitis”,
Proceedings of ICSLP96, Philadelphia, pp. 749 –
752 , 1996
[4] M.N. Viera, F.R. McInnes, M.A. Jack “Robust
F0 andJitter estimation in the Pathological voices “,
Proceedings of ICSLP96, Philadelphia, pp.745 –
748, 1996.
[5] J.Nayak, P.S.Bhat “Classification and analysis
of speech abnormalities”, ITBM-RBM 26 (2005)
319-327.
[6] S. Mallat, “A Theory for multiresolution signal
5 6 7
Trans. Pattern Analysis and Machine Intelligence.
Vol. 11. No. 7 pp674-693 July 1989.
[7] B. Boyanov, S.Hadjitodorov: “Acoustic analysis
of pathological voices: a voice analysis system for
screening of laryngeal diseases”, Proc. IEEE
Engineering in Medical and Biology, (1997), vol.
16, no. 4, 74-82.
[8] J.J. Jiang,Yu Zhang: “Nonlinear dynamic
analysis of speech from pathological subjects”,
Proc IEEE Electronics Letters, March (2002),
vol.38, no.6.
[9] P.Yu, M.Ouaknine, J.Revis, and A.Giovanni,
“Objective Voice Analysis for Dysphonic Patients:
A Multiparametric Protocol Including Acoustic and
Aerodynamic Measurements”, Journal of Voice
Vol. 15, No. 4, pp. 529–542 © 2001 The Voice
Foundation
[10] J.Wang, Jo.Cheolwoo, “Performance of
Gaussian Mixture Model as a classifier for
Pathological Voice”, proceeding of the ASST in
Auckland 2006, pp 165-169.
[11] J.Kortelainen, K.Noponen, « Neural
networks », Intelligent Systems 2005
[12] S.Lotfi, C.Adnène, “A Speech Processing
Interface for Analysis of Pathological Voices”, in
proceeding of ICTTA conference, Damascus 2006.
[13] S.Lotfi, B.Haythem, C.Adnène, “ Interface
d’analyse vocale a l’identification de certaines
pathologies d’origine neurologique et vocale”, in
proceeding of JTM conference, Tunis 2007.
[14] A.CHÉRIF, « Pitch detection and formant
extraction of Arabic speech processing » Journal of
applied acoustics, January 2001
[15] A.M. Gaouda, M. Salama, A. Chikhani, and
M. Sultan, “Application of wavelet analysis for
monitoring dynamic performance in industrial
plants,” North American Power Symposium, Oct.
1997, Laramie, Wyoming.
-6-