Journal of Signal and Information Processing, 2011, 2, 336-347
doi:10.4236/jsip.2011.24048 Published Online November 2011 (http://www.SciRP.org/journal/jsip)
1
A New Method of Voiced/Unvoiced Classification
Based on Clustering
Mojtaba Radmard, Mahdi Hadavi, Mohammad Mahdi Nayebi
Department of Electr ical Engineering, Sharif Univer sity of Techno logy, Tehran , Iran.
Received June 2nd, 2011; revised October 14th, 2011 ; accep ted October 23rd, 2011.
ABSTRACT
In this paper, a new method for making v/uv decision is developed which uses a multi-feature v/uv classification algo-
rithm based on the analysis of cepstral peak, zero crossing rate, and autocorrelation function (ACF) peak of short-time
segments of the speech signal by using some clustering methods. This v/uv classifier achieved excellent results for iden-
tification of voiced and unvoiced segments of speech.
Keywords: Speech, Voiced, Unvoiced, Clustering, Cepstrum, Autocorrelation, Zero Crossing
1. Introduction
The voiced/unvoieced decision is critical in many speech
analysis/synthesis systems because it is essential to know
whether the speech production system involves vibration
of the vocal cords [1-4]. This decision is required for many
applications, including modeling for analysis/synthesis,
detection of model changes for segmentation p ur p oses a nd
signal characteriza tion for ind exing and recognition appli-
cations [1]. The periodicity of this vibration makes the
voiced segments periodic and so distinguishable from the
noisy-like unvoiced segments [5]. Since the speech signals
are quasi-periodic [6-9], making the decision gets hard.
Other difficulties are represented in [3,10].
Common methods extract a feature from speech seg-
ments and make the v/uv decision according to whether
the value of the feature is above or b elow a pre-determined
threshold. The feature can be the cepstral peak [6,11],
some mel-frequency cepstral coefficients [12,13], energy
of the segments [3,14], zero-crossing rate [3,14], the a uto-
correlation function peak [15,16], or harmonic to noise
ratio in the sinusoidal model of speech signal [17]. Since
each feature has its own disadvanta ges, ne w metho ds tend
to use a c ombin ation of features for v/uv decision [2,3] and
since the value of these feat ures are di fferent for variety o f
speeches, adaptive thresholding have been used in most of
the methods.
Different methods have been used in the field of multi-
feature voicing decision. Atal and Rabiner [3] clustered
the segments into two major groups based on a weighted
Euclidian distance in the feature vector space while the
weight was estimated according to some statistical prop-
erties and the features were considered gaussian distrib-
uted in each cluster. Siegel [18] used a non-statistical
nonparametric classifier to make v/uv decision. In this
method no assumptions are made about the distribution
of the features, and training focuses on patterns near the
boundry between two regions in the feature space, rather
than using statistics to describe each class. Siegel and
Bessey [19] tried to develop the mentioned methods by
using linear discrimination in the feature vector space.
The use of two or more feature s in the voici ng decision
tended to the methods which do not consider the features
as a vector and use the best feature for each frame like the
work in [20] . The work presente d here is in this categor y.
In this work, which is preliminaril y presented in [21], we
use implicitly two thresholds for each feature and make
the decision for each frame based on only one of the fea-
tures that performs better than other features. This paper
is organized as follows. The description of the suggested
algorithm is given in Section 2. Some discussions about
the new method are presented in Section 3. In Section 4,
the results of the algorithm are represented. Section 5
discusses the disadvantages of former methods. Finally,
the conclusion is given in Section 6.
2. The Proposed Algorithm
In this section we will describe our method for V/UV de-
cision. Here we use three features which are the cepstral
peak, autocorrelatio n function (ACF) peak a nd zero cross-
ing rate. The speech signal, sampled at 8 kHz, is analyzed
at 10 ms intervals using a 40 ms Hamming window. Then

A New Method of Voiced/Unvoiced Classification Based on Clustering
337
the following features are extracted and analyzed.
1) Cepstral peaks: The cepstrum, defined as the real
part of the inverse Fourier transform of the log-power
spectrum, has a strong peak corresponding to the pitch
period of the voiced speech segment being analyzed [22].
Here we use a primary normalization to have a fair deci-
sion fo r all of the frames (i ncludin g high ene rgy and lo w
energy frames). A 512-point fast Fourier transform (FFT)
is used and the peak picking scheme is to determine the
cepstral peak in the interval [2.5 - 15 ms], corresponding
to pitch frequencies between 60 - 400 Hz. Since the cep-
stral peaks decrease in amplitude with increasing que-
frency, a linear cepstral weight is applied over the 2.5 to
15 ms. The linear cepstral weighting with range of one to
five was fo und e mpiric ally b y usin g per iodic pulse train s
with varying periods as the input to the program.
2) Zero crossing rate: the method in this part is the
well known met ho d using the formula (1).
( )()
1
=1
=1
N
ii i
n
ZCRsgn xnsgn xn
−
−−
 
 
∑
(1)
It is kno wn than tha t the ZCR of an unvoice d segme nt
is much mor e tha n that of a voiced segment.
3) Auto-correlation function peaks: As we know, the
speech signal is periodic for voiced segments. So, we
make the V/UV decision based on finding a high peak in
this function. But since in this algorithm, this function
should behave fairly for different segments, it should be
normalized, like the cepstral peaks. But, since the speech
signal is originally quasi-periodic at voiced segments and
the noise of the environment is added, the voiced seg-
ments are not precisely periodic. This non-periodicity
emerges in the segments with low energy (because SNR
falls down) or in high frequencies (since noise is masked
in low frequencies but can be destructive in high frequent-
cies). To eliminate the first effect, we use the method of
Center clipping [10], with the clipper's amplitude of 1/3
of the maximum of the absolute of the signal's amplitude.
To eliminate the second one, we use a band-pass filter
with the cut-off frequencies of 20 and 900 Hz.
After determining how each feature is extracted, we go
back to the algorithm. Here the V/UV decision is made in
this way: each of the information groups (that obtained
from features: Cepstral peaks, ACF peaks and zero cro-
ssing rate) is clustered into three clusters by K-Means
algorithm [23]. So after clustering, we have three clusters
for each feature: the first cluster contains the frames in
which the related feature has low values, the second one
contains the frames in which the related feature has av-
erage values and the third one contains the frames in
which the related feature has high values. For example,
the frames in the first cluster of zero crossing have low
ZCR and so, are very likely to be voiced, despite the
third cluster that are very likely to be unvoiced and about
the second cluster we cannot conclude yet. But when we
consider the three features simultaneously, we can decide
about almost all of the frames. It is practically observed
that very little fra mes are found to be in the second clus-
ter for all three features (about 4% of all the frames). We
make the V/UV decision for these frames by clustering
them into two clusters, based on autocorrelation funct ion
(since it works better than the other features as we will
see). Now the only thing remained to do is to decide
what to do about the frames that are voiced in a feature
and unvoiced in another feature. Here, priority gets im-
portant. It means we decide the V/UV of a frame after
giving each cluster a priority. The six rules that we
choose to determine the priorities are as below:
If a frame belongs to the first cluster of zero-crossing
rate, it is voiced .
If a frame belongs to the third cluster of zero-crossing
rate, it is unvoiced.
If a frame belongs to the first cluster of cepstral peaks,
it is unvoiced .
If a frame belongs to the third cluster of cepstral peaks,
it is voiced.
If a frame belongs to the first cluster of ACF peaks, it
is voiced.
If a frame belongs to the third c luster of ACF peaks, it
is unvoiced.
How these rules are given priorities, is described be-
low:
First, for each rule we calculated the error probability,
and the one with the least error probability was chosen as
the first priority. Then, to choose the second priority, we
calculated t he conditional err or prob ability for the rest o f
the rules, on the condition that the first priority is defined
and classifies some frames as voiced or unvoiced (based
on which rule is considered as the first priority). For
these tests, we used some TIMIT files. We continued in
this way until all p riorities are d efined. The results are as
below:
The 1st p r iority: the third clust e r of ACF peaks.
The 2nd priority: the third cluster of cepstral peaks.
The 3rd priority: the third cluster of zero-crossing rate.
The 4th pr iority: the first clust e r of zero-crossing rate.
The 5th pr iority: the first clust e r of ACF peaks.
The 6th pr iority: the first clust e r of cepstral pe a ks.
Also you will see the complete results with the error
prob abilities in t he simulation and e valuation sectio n.
To improve the performance of the clustering algo-
rithm, we used a limiter for each feature. The reason is
that some frames have large values (e.g. ACF peaks) and
this causes the cluster ing algorithm to consider them as a
separate cluster in the third cluster. The upper bound for
each feature is chosen proportional to the mean of that

A New Method of Voiced/Unvoiced Classification Based on Clustering
338
feature in all frames. Also, in order to choose a better
upper bound, we eliminated the silence of the start and
end of the speech. At last, a median filter of order 5 is
found empirically to work well for the resulting V/UV
estimates. The block diagram of the proposed algorithm
is depicted in Fig ur e 1.
3. Discussion
The reason of using three features with three clusters is
that each feature will perform well and accurately, just
when the value of that feature is very high or very low.
So the use of three clusters will help us in this matter.
Also to complete the decision, it is necessary to use more
than one feature. In this case each feature will correct
some of the other features’ mistakes, because each fea-
ture's base is different from the other one.
Also three clusters can be considered as using two
threshold s (similar to double threshol ding in the detection
topics [24]). So it is obvious that using more than one
feature with e ach having two thr esholds will pe rform bet-
ter than using some features with one threshold and that
will work better than using one feature with one thresho ld,
which is usually used to make a V/U V decision.
The deficiencies of different methods and different
features in V/UV decision are discussed below and we
will show the deficiencies and faults of each feature with
some practical samples.
In the cepstral domain, the considerations and tests'
results show that the cepstral peak does not perform well
when the signal has limited bandwidth. Inspite of that,
whe n the signal has high fr equenc y coeffici ents, the cep-
stral peak is a good indicator of being voiced/ unvoiced.
The empirical results that show this can be seen in Fig-
ures 2 and 3.
Figur e 1. The block diagram of the proposed algorithm.

A New Method of Voiced/Unvoiced Classification Based on Clustering
339
number of samples
(a)
number of samples
(b)
number of samples
(c)
Figur e 2. Testing the cepstral feature. (a) The signal with limited bandwidth; (b) The spectral domain; (c) The cep s t ral domain.

A New Method of Voiced/Unvoiced Classification Based on Clustering
340
number of samples
(a)
number of samples
(b)
number of samples
(c)
Figure 3. Testing the cepstral feature. (a) The signal with high frequency coefficients; (b) The spectral domain; (c) The cep-
stral domain.
Also about the ACF, the effects of the vocal source
and vocal tract are convolved with each other in the au-
tocorrelation functions and this results in broad peaks
and in some cases multiple peaks in the autocorrelation
function [22]. Furthermore, the considerations and tests’
results show that the ACF peak does not perform well
whe n the fu nd a me nta l fre q uen c y and the first for mant are
near each other. This fact can be easily seen in Figure 4.

A New Method of Voiced/Unvoiced Classification Based on Clustering
341
number of samples
(a)
number of samples
(b)
number of samples
(c)
Figure 4. Testing the ACF feature. (a) A speec h signal wi th the funda mental f requenc y near the first for mant; (b) The s pec-
tral d omain; (c) The ACF.

A New Method of Voiced/Unvoiced Classification Based on Clustering
342
Besides, there are segments in speech that are not pe-
riodic but are similar to noise. Inspite of that, their ZCR
is low. As an example, the silence segments, between
speech, have very low energy but if classified based on
their ZCR, they are mistakenly marked voiced. Compar-
ing Figures 5(a) and (b) shows thi s fa ct.
In clustering we cluster the values of each feature in
all frames into 3 clusters: low, average and high values.
In fact, when speech is uttered by a specific person, each
feature’s values in the voiced segments are very similar
to each other. This is true about the unvoiced segments
too and this is the base of the clustering method, which
performs very we ll through the tests.
4. Simulations’ Results
In this section we show the simulations’ results and dis-
cuss about the quality of the proposed algorithm. 821
frames of speech, that were taken from TIMIT, were
tested. To calculate the error probability of each of the
rules (the six rules described above with considering the
priorities we defined), we counted the number of frames
that were classified as voiced or unvoiced in each rule
(each priority) based on the priorities we determined.
Then we counted the number of frames, which were
wrongly classified. The frames were labeled visually by
looking at their time domain shape and their frequency
domain spectrum. The results are depicted in Table 1.
Totally the error for voiced segments was 4.8% and the
error for unvoiced segments was 1.1%.
In the above table etc means the number of frames that
are clustered to the second cluster for all three features
and are classified as voiced or unvoiced by the clustering
number of samples
(a)
number of samples
(b)
Figure 5. Testing the ZCR feature. (a) A scilence segment with low energy and low ZCR; (b) An unvoiced segment with high ZCR.

A New Method of Voiced/Unvoiced Classification Based on Clustering
343
Table 1. Simulations ’ result s.
auto-3
(1st priority) ceps-3 (2nd priority) zero-3 (3rd priority) zero-1 (4th pr io rity) auto-1 (5th pr io rity) ceps-1 (6th pr io rity) etc
The number of
frames identified
V or UV 248 46 158 124 142 69 34
The number of
frames wrongly
identified
0 0 1 2 7 10 3
method described in section two (clustering to two clus-
ters based on autocorrelation function). To better under-
stand the above table, for example, it shows that 46 fr-
ames were clustered in the third cluster of cepstrum (se-
cond priority) but not in the third cluster of the autocor-
relation function (first priority). So they were classified
as voiced and according to the table none of them were
misclassified.
5. Disadvantages of the Cepstrum-Based
Voicing Detector
One of the most common methods for extracting pitch
period is to determine the place of the peak in the cep-
stral domain [10]. Furthermore, the cepstral domain in-
formation is used to extract other acoustic parameters.
One of these important parameters is “voicing” which is
extracted based on the peak value in the cepstral domain
by using different methods. But the first problem is that
the peak value in the cepstral domain depends on its
place on the cepstral axis. So when the pitch period gets
larger, the peak value descends with rate
1n
. T he solu-
tion is to multip ly a ramp function in the cepstra l domain.
More details are presented in [20]. More experiments
have shown that the cepstral method has other deficien-
cies. It means that for some voiced frames, although
there is a distinguishable peak in the cepstral domain, it
does not have sufficient value in comparison with the
threshold.
The frames for which cepstrum method cannot per-
form well can be divided into two categories. For each
category a sample frame from TIMIT directory is ana-
lyzed which shows the deficiency of the method obvi-
ously. Note that both autocorrelation and cepstrum me-
thods are used after the energy-normalization of the
frame.
In the categories, in order to prove our claim about all
the frames in the gro up, we mo del the cate gor y with some
known mathematical functions such as “sine” and “sinc”.
The reason that the function can simulate nearly all the
frames in the category is discussed in related sections.
The first category contains the vowels which are
band-limited in the spectral domain. In this case the cep-
stral peak value is small and this leads to wrong v/uv
classification in the cepstrum-based methods. To prove
our claim, we consider a periodic “sinc” function as the
input of the cepstrum method. Note that the spectral peak
values for this input are the same as each other and if we
want to have a spectral shape similar to the frames in this
category, we need to multiply this function with some
formant-like function in the spectral domain. As it is
known, because of the logarithmic property of cepstrum,
this multiplication in the spectral domain will result in
addition in the cepstral domain and since the periodicity
information is in the periodic sinc function, the result of
cepstrum method for the periodic sinc function will be
similar to the result of applying the method to any frame
in thi s category. Figuers 6 and 7 shows the results of this
application to two different sincs, one with limited band-
width and the other with high frequency coefficients.
It can easily be seen that by increasing the bandwidth
of the signal the value of the cepstrum feature has in-
creased from 3.69 to 6.
For more support of our claim we have plotted the
value of the cepstrum feature by increasing the band-
width for a periodic sinc function. The result is depicted
in Figure 8. It can be seen that this value increases as the
bandwidth increases, meaning that the cepstrum performs
better.
The similar results for applying the method to a prac-
tical frame of a vowel (/i:/ like in sheet) are shown in
Figure 9.
The reason can be explained mathematically for a pe-
riodic sinc function (which is an indicator of a band li-
mited signal) as this:
The cepstrum is evaluated from the Equation (2) [10]:
( )( ) { }
1
= log
s
c nsn
−

(2)
As we know if
( )
sn
is a periodic sinc, the fft of its
absolute value will be t he mul tip lication o f a p ulse with a
delta train. Then, its log will also contain some deltas
(the deltas within the pulse width). The larger the sinc’s
BW is (in other words, the sharper the sinc is) the more
deltas will be included in the pulse width (and therefore
more deltas we will have at the output of the fft). Consi-
dering that these deltas show the periodicity of the origi-
nal wa vefo r m (si nc), by i ncre asin g the BW , the o utp ut o f
the ifft (in the cepstrum equation) will have larger value
at the pitch.

A New Method of Voiced/Unvoiced Classification Based on Clustering
344
The second category contains frames related to nasals
such as / n / and / m / which must be labeled voiced in a
correct voicing decision. But the theoretical and practical
results show that their cepstral peak values are so small.
For modeling nasals to study them, we choose a sine
wave, which is a good indicator of nasals.
A theoretical conclusion similar to the one in the first
category can be made here. The results of applying this
explicit waveform can be seen in Figure 10.
The results of applying the method to a practical frame
(/n/ in background) are a l s o shown in Figure 11.
As can be seen for both vowels and nasals, cepstrum
based methods do not perform well to extract the pa-
rameters of the speech segment, such as voicing and
pitch. That's why we do not rely on just one feature. Also
we add a third group for each feature (besides the voiced
and unvoiced groups), so that if that feature is weak in
making the dec i s ion, we can go t hroug h other features.
6. Conclusions
We have presented a new approach of detecting voiced
and unvoiced speech. The main advantage of this clus-
tering-based method is getting rid of determining a thre-
shold. So it is highly speaker independent. Also, the use
of three features has enabled the method to make a better
decision about the segments, in which one feature does
not indicate voicing well. Besides, clustering into three
clusters, or implicitly, double thresholding, helps us to
make the v/uv decision more certainly. Despite the sim-
plicity of the algorithm, the results have shown a satis-
factory performance in comparison with more compli-
cated methods.
number of samples
number of samples
(a) (b )
number of samples
(c)
Fig ure 6. Testi ng the si nc funct ion for t he first categ ory. (a) A sample sinc function with limited bandwidth; (b) T he spectral
doma in; (c) The cepst ral domain, peak to average = 3.69.

A New Method of Voiced/Unvoiced Classification Based on Clustering
345
number of samples
number of samples
(a) (b)
number of samples
(c)
Figure 7 . Testing the sinc fu nction for the first c ategory. (a) A sample si nc function w ith high freq uency coeffici ents; (b) The
spectral domain; (c) The cepst ral domain, peak to average = 6 .
number of samples
Figure 8. Cepstral feature for the s inc f unction when incre a sing BW.

A New Method of Voiced/Unvoiced Classification Based on Clustering
346
number of samples
number of samples
(a) (b)
Figure 9 . Testing a real sp eech fra me f or the first category. (a) The vowel / i:/ as in sheet; (b) The cepstral domain.
number of samples
number of samples
(a) (b)
Figure 1 0. Testing a sample sine functi o n. (a) The sample sine functi on; (b) The cepstra l do main.
number of samples
number of samples
(a) (b)
Figure 1 1. Testing a real speech fra me f or the second category. (a) The nasal /n/; (b) The cepstral domain.

A New Method of Voiced/Unvoiced Classification Based on Clustering
347
REFERENCES
[1] E. Fisher, J. Tabrikian and S. Dubnov, “Generalized Li-
kelihood Ratio Test for Voiced-Unvoiced Decision in
Noisy Speech Using the Harmonic Model,” IEEE Trans-
actions on Audio, Speech, and Language Processing, Vol.
14, No. 2, 2006, pp. 502-510.
doi:10.1109/TSA.2005.857806
[2] Y. Qi and B. R. Hunt, “Voiced-Unvoiced-Silence Classi-
fications of Speech Using Hybrid Features and a Network
Classifier,” IEEE Transactions on Speech and Audio
Processing , Vol. 1, No. 2, 2002, pp. 250-255.
doi:10.1109/89.222883
[3] B. Atal an d L. Rab in er, “A P atter n R ecogn ition Approach
to Voicedunvoi ced-Silence Classification with Applica-
tions to Speech Recognition,” IEEE Transactions on Ac-
oustics, Speech and Signal Processing, Vol. 24, No. 3,
2003, pp. 201-212. doi:10.1109/TASSP.1976.1162800
[4] F. Y. Qi and C. C. Bao, “A Method for
Voiced/Unvoiced/Silence Classification of Speech with
Noise Using SVM,” Acta Electronica Sinica, Vol. 34, No.
4, 2006, pp. 605-611.
[5] P. Jancovic and M. Kokuer, “Estimation of Voicing-
Character of Speech Spectra Based on Spectral Shape,”
IEEE Signal Processing Letters, Vol. 14, No. 1, 2006 , pp .
66-69. doi:10.1109/LSP.2006.881517
[6] B. Atal and M. Schroeder, “Predictive Coding of Speech
Signals and Subjective Error Criteria,” IEEE Transac-
tions on Acoustics, Speech and Signal Processing, Vol.
27, No. 3, 2003, pp. 247-254.
doi:10.1109/TASSP.1979.1163237
[7] L. Hui, B. Dai and L. Wei, “A Pitch Detection Algorithm
Based on AMDF and ACF,” 2006 IEEE International
Conference on Acoustics, Speech and Signal Processing,
Toulouse, 14-19 May 2006.
[8] P. A. Naylor, A. Kounoudes, J. Gudnason and M. Brookes,
“Estimation of Glottal Closure Instants in Voiced Speech
Using th e DYPSA Algorithm,” IEEE Transactions on Au-
dio, Speech and Language Processing, Vol. 15, No. 1,
2007, pp. 34-43. doi:10.1109/TASL.2006.876878
[9] A. V. Oppenheim, “Speech Spectrograms Using the Fast
Fourier Transform,” IEEE Spectrum, Vol. 7, No. 8, 2009,
pp. 57-62. doi:10.1109/MSPEC.1970.5213512
[10] J. R. Deller, J. G. P ro ak is and J. H. L. Hansen , “Discrete-
Time Processing of Speech Signals,” 2nd Edition, IEEE
Press , New York, 2000.
[11] Z. D. Zhao, X. M. Hu and J. F. Ti an, “An Effecti ve Pitch
Detection Method for Speech Signals with Low Sig-
nal-to-Noise Ratio,” International Conference on Ma-
chine Learning and Cybernetics, Vol. 5, 2008, pp. 2775-
2778,.
[12] S. Imai, “Cepstral Analysis Synthesis on the Mel Fre-
quen cy Scale,” IEEE International Conference on Acous-
tics, Speech, and Signal Processing, Vol. 8, 2003, pp.
93-96.
[13] J. K. Shah, A. N. Iyer, B. Y. Smolenski and R. E. Yan-
torno, “Robust Voiced/Unvoiced Classification Using
Novel Features and Gaussian Mixture Model,” IEEE In-
ternational Conference on Acoustics, Speech, and Signal
Processing , Philadelphia, 2004, pp. 17-21.
[14] R. G. Bachu, S. Kopparthi, B. Adapa and B. D. Barkana,
“Separation of Voiced and Unvoiced Using Zero Cro ss-
ing Rate an d Energy of the S peech Signal,” A merica n S o-
ciety for Engineering Education (ASEE) Zone Conference
Proceedings, 2008, pp. 1-7.
[15] L. Rabiner, “On the Use of Autocorrelation Analysis for
Pitch Detection,” IEEE Transactions on Acoustics, Speech
and Signal Processing, Vol. 25, No. 1, 2003. pp. 24-33.
doi:10.1109/TASSP.1977.1162905
[16] M. S. Rahman and T. Shimamura, “Pitch Determination
Using Autocorrelation Function in Spectral Domain,”
Eleventh Annual Conference of the International Speech
Communication Association, Makuhar i, 2010, pp. 653-
656.
[17] R. J. McAulay and T. F. Quatieri, “Pitch Estimation and
Voi cing Detectio n Based on a Sinusoidal Speech Model,”
International Conference on Acoustics, Speech, and Sig-
nal Proc e s s ing, Vol. 1, 1990, pp. 249-252.
doi:10.1109/ICASSP.199 0.115585
[18] L. Siegel, “A Procedure for Using Pattern Classification
Techniques to Obtain a Vo iced /Un voi ced Cl as si fier ,” IEEE
Transactions on Acoustics, Speech and Signal Processing,
Vol. 27, No. 1, 2003, pp. 83-89.
doi:10.1109/TASSP.1979.1163186
[19] L. Siegel and A. Bessey, “Voiced/Unvoiced/Mixed Exci-
tation Classification of Speech,” IEEE Transactions on
Acou stics, Speech and Signal Processing, Vol. 30, No. 3,
2003, pp. 451-460. doi:10.1109/TASSP.1982.1163910
[20] S. Ahmadi and A. S. Spanias, “Cepstrum-Based Pitch
Detection Using a New Statistical V/UV Classification
Algorithm,” IEEE Transactions on Speech and Audio
Processing, Vol. 7, No. 3, 2002, pp. 333-338.
doi:10.1109/89.759042
[21] M. Radmard, M. Hadavi, S. Ghaemmaghami and M. M.
Naye bi, “Cluster ing Based Voiced/Un voiced Decis ion for
Speech Signals,” Signal Processing Symposium (SPS),
Poland, 2011.
[22] A. M. Noll, “Clipstrum Pitch Determination,” The Jour-
nal of the Acoustical Society of America, Vol. 44, No. 6,
1968, pp. 1585-1591. doi:10.1121/1.1911300
[23] J. A. Hartigan and M. A. Wong, “A K-Means Clustering
Algorithm,” Journal of the Royal Statistical Society. Se-
ries C, Vol. 28, No. 1, 1979, pp. 100-108.
[24] H. V. Poor, “An Introduction to Signal Detection and
Estimation,” Springer, Berlin, 1994.