CC BY
8
НЕИРОШФОРМАТИКА ТА 1НТЕЛЕКТУАЛЬН1 СИСТЕМИ
НЕИРОИНФОРМАТИКА И ИНТЕЛЛЕКТУАЛЬНЫЕ СИСТЕМЫ
NEUROINFORMATICS _AND INTELLIGENT SYSTEMS_
UDC 681.327.12
ANALYSIS OF THE AUTOMATED SPEAKER RECOGNITION SYSTEM OF CRITICAL USE OPERATION RESULTS
Bisikalo O. V. - Dr. Sc., Professor, Dean of Faculty of Computer Systems and Automation, Vinnytsia National Technical University, Vinnytsia, Ukraine.
Kovtun V. V. - PhD, Associate Professor, Computer Control Systems Department, Vinnitsia National Technical University, Vinnitsa, Ukraine.
Yukhimchuk M. S. - PhD, Associate Professor, Computer Control Systems Department, Vinnitsia National Technical University, Vinnitsa, Ukraine.
Voytyuk I. F. - PhD, Associate Professor, Computer Science Department, Ternopil National Economic University, Ternopil, Ukraine.
ABSTRACT
Context. The article summarizes the statistical learning theory to evaluate the long-term operation results of the automated speaker recognition system of critical use (ASRSCU) taking into account the features of the system's operation object and the structural specificity of such a class of recognition systems.
Objective. The goal of the represented work is the development of a complex set of methods for the ASRSCU's quality parameters stabilization during its long-term operation.
Method. The article formulated set of methods for the ASRSCU's operational risks estimation of its long-term operation. In particular, the dependence of the risk of an incorrect speaker recognition on the features space dimension is described. Based on the formulated measure of informativity, obtained a set of methods to analyze the training sample to identify examples that lead to increased risk. The influence of the phenomenon of the drift of the speech signal parameters on the quality indicators of the ASRSCU is described analytically. An estimation of the operation duration of the ASRSCU, during which it is impractical to re-train its the classifier, is carried out. Recommendations for choosing an optimal ASRSCU's classifier are formulated from the position of its complexity minimization, taking into account the risks of the ASRSCU's long-term operation and the possibility of re-training.
Results. Represented in the article theoretical results are verified by the DET-curves experiments data, which summarize the information from long-term experiments with the ASRSCU, in which, during the features space configuration were taken into account the features based on the power normalized cepstral coefficients based and the features based on the spectral-temporal receptive fields theory. Within the framework of the created theoretical concept, an estimation of the influence of the features space configuration and the type and complexity of the classifier on the stability of the ASRSCU's quality parameters during its long-term operation has been carried out.
Conclusions. For the first time the theoretically analyzed the problem of average risk minimization by empirical operation results of a ASRSCU, where, unlike existing approaches, non-stationary input data with the drift of individual speech signals features and the characteristic parameters of the recognition system classifier were taken into account, which allowed to estimate the risk's confidence interval for conditions for re-training sessions.
KEYWORDS: automated speaker recognition system of critical use, experiment planning theory, factor analysis, statistical learning theory.
ABBREVIATIONS NOMENCLATURE
ASRSCU is automated speaker recognition system of A is a impulsive variable; critical use; Q is a parameter of temporal impulse filter
CNN is a convolution neural network; responses;
STRF is a spectral-temporal receptive fields; a is a set of parameters of the classifier;
VAD is a voice activity detection; e is a accuracy of classifier training;
8 is a phase of spectral impulse filter feedback; k1 is a weight constant;
k2 is a time constant;
k3 is a constant, which set the CNN parameters; k4 is a constant, which set the CNN parameters; k5 is a constant, which set the CNN parameters; k6 is a constant, which set the CNN parameters; k7 is a training rank;
is a set of n independent, equally distributed random variables;
p is a probability that, at least in one of the N
functions Q(z,ak), k = 1,...,N , the upper limit of the risk Rmax exceeds Rm ;
t is a mathematical expectation of the interval between sequences of re-training;
9 is a phase of spectral impulse filter feedback;
..) is a function of estimation of the drift degree of the input data;
ra is a density parameter of the spectral impulse response filters;
{((, yi)} is a the element of the training sample, X and Y are the set of empirical input and output data of the system;
1 -8 is a reliability of the classifier's training;
b" is a offset for n -i source map,
d is a degree of complexity of the training algorithm;
F1(t, ra) is a SNRF-individual feature;
F2 ((, ra) is a SNRF-individual feature;
F3 ((,k) is a SNRF-individual feature;
f is a frequency;
G(l) is a measure of the learning process complexity;
g (x) is a classification function;
h is a measure of Vapnik-Chervonenkis;
h is a Gilbert transformation;
hs is a spectral impulse response;
hsc is a spectral scale factor;
hT temporal impulse response;
hrt is a temporal scale factor;
h((, f) is a impulse response of each filter in bank;
H is a class of hypotheses of indicator functions g ;
i is a iterator;
I (...) is a informative measure of rejection of empirical data from educational; j is a iterator;
Lq (h) is a loss function, which describes the average difference between a random variable Y and h(X ); Lq (h) is a empirical risk;
l is a basic length of the training sample; lm is a expected average durability of the recognition system exploitation term without re-training; m is a iterator;
M"-1 is a number of entrance maps; n is a iterator;
Nra is a number of elements hcc; Nk dimension of the feature F3 (t, k); k is a iterator;
O is a degree of sufficiency of the training sample; Pa is a probability of the first kind errors of ASRSCU;
Pp is a probability of the second kind errors of ASRSCU;
P is a unknown probabilities distribution; P is a empirical risks by sampling Sl subset A ; P2 is a empirical risk by sampling S2 subset A ; Pl (• • •) is a function of estimating the marginal size of the initial sample;
Pr is a probability of a situation, when testing the system on a plurality m of elements, the empirical risk Re will exceed the threshold value of risk Rr;
Q (...) is a set of corresponding indicator functions;
R assessment of empirical risk Re;
R(a) is a functional risk;
Re (a) is a functional of empirical risk;
Rg is a risk of incorrect classification when training a
classifier on a sample from which elements were removed at the initial sample length, estimated by the error rate;
Rm is a threshold value of the risk, determined by the testing results;
Rmax is a upper limit of risk;
STRF (t, f, Q, ra, 9,8) is a operation of allocation of SNRF features; t is a time;
w",m is a convolution core between m -input and n -output source maps;
X-"" is a input feature map m of a layer u ; y(t, f ) is a SNRF spectrogram; yA (t, f) is a model of hair cells work; yc (t, f) is a affine wavelet transform of the speech signal frame c(t);
yLIN (t, f) is a model of the lateral inhibitory network;
y" is a output feature map n of a layer u ; Z = {z1 = (X, yi), z2,..., zi} is a set of empirical data.
INTRODUCTION
The automated speaker recognition system of critical use [1], as well as all speaker recognition systems, performs the speaker's person recognition by analyzing the individual attributes isolated from the phonogram with the recording of the speech signal. Of course, the speaker is characterized by the pronunciation variability, due both to internal and external factors. To the internal speech variability factors, we will relate the style, tempo and volume of speech. External speech variability factors are characterized by the type and level of noise in the acoustic and hardware channels of the speech signal propagation, as well as distorted perception of the speech signal due to the reverberation of the speaker's spatial surroundings. Also highlighted such high-level individual characteristics of speech as dialect and speech style, which manifests itself in the acoustic characteristics of the speech signal and the tempo of speech. To establish in ASRSCU the potential for distinguishing internal variability factors, taking into account the high-level individual characteristics and resistance to external variability factors, can be used in the systematic approach to forming the features space, the selection and parameterization of classifier, the formation of a training sample and the regulation of the training process. There are other "extreme" volatility sources of the amplitude-frequency characteristics of the speech signal due to the state of the speaker's health, the acoustic parameters and the geometry of the room where the system is operate, the parameters and the location of the microphones. However, these types of variability are so significantly distorting the meaning of informative speech recognition features that they are reasonably easily identifiable and taken into account when deciding on the result of a speech recognition session, taking into account the degree of distortion and the scope of use and the type of the recognition system.
However, the study [2] showed that during prolonged use of the speaker recognition system, the speech signal parameters drift is due to simple normal physiological processes in the articulatory apparatus of the human, as a result of which the time difference between the training session of the ASRSCU classifier and the recognition session can significantly affect the quality system performance. Consequently, the possible critical use of the speaker recognition system necessarily requires the study of the influence of the operating time on the qualitative performance indicators of the recognition system in order to stabilize them.
The object of study is the individual features of the process of human speech activity and the process of hearing perception of speech signals by a human being and their analysis by the auditory cerebral cortex.
The subject of study is the methods of the pattern recognition theory for the modeling of the recognition system, the methods of the statistical training theory for the analysis of the risks arising from the long-term use of the recognition system, the methods of the neural networks theory for the implementation of the optimal
classifier for the recognition system and methods of spectral-temporal receptive fields to describe the process of perception speech signals to the acoustic cerebral cortex of a human.
The purpose of the work is to estimate the risks of long-term operation of the ASRSCU and to propose measures to reduce them.
1 PROBLEM STATEMENT
Let there exist an abstract teacher who offers the ASRSCU's classifier a finite set of examples of an unknown indicator function g over the domain X. On
the basis of a sequence of examples {(Xj,yj)}, i = 1,...,l, Xj e X, yi e Y, where yi = g (xi), 1 < i < m , and X <z R , Y = {0,1}, randomly selected in accordance with an unknown distribution of probabilities above X, it is necessary to train a classifier with given accuracy e > 0 and reliability 1 - 8. Suppose the existence of an unknown probabilities distribution P over X x Y , a classifying function g(x) = _e(y|X = x) and a limited
trainning sample with examples {((,yi)}, i = 1,.,l, where (xj, yj) are taken independently, respectively P over X x Y . The classifying function belongs to the hypothesis class H , which does not necessarily contain g . Determine the loss function Lq (h), which describes
the average difference between the random variable Y
and h(X) : Lq(h) = ( E\Y - h(X)
and the empirical
risk
Lq (h) : Lq (h) = l-1 £ (y - h(xi )q } , where the
i=1
q > 1 averaging takes place in accordance with P . If we take into account Y = g (X), then the loss function will
that is Lq the norm
take the
form Lq (h) = lg - h\Lq (P )
over P . Suppose the existence of a classifier training algorithm, a coincidence l (e,8) for examples of a training sample, for any objective function g and any probability distribution P on X , if the hypothesis is h e H fulfilled Lq (h)< Lg q +e with probability greater
than 1 - S , where
Lg ,q = ¿fig - % (P )
- the loss of
the optimal hypothesis H . This assumption suggests that it is possible to minimize the empirical risk by using the
classifier training algorithm, which h is the result of a
choice h with H a minimum value Lq (h), and
formulate the purpose of the training procedure as the choice of an element from H , which minimizes the
generalization error R = J H (z )dP(z ) based on empirical data Z = {z1 = (xj-,y{),z2,...,}. Next in the article, we will develop the concept formulated above in direction of
q
identifying the relationship between the performance indicators of the ASRSCU and the training process parameters and the recognition system classifier parameters for its long-term operation.
2 REVIEW OF THE LITERATURE
In the theory of pattern recognition, one of the applied applications of which is the speaker recognition systems, one of the central problems is to minimize the average risk based on the analysis of empirical data, which developed into the theory of statistical learning [3]. In this theory, many complex problems are investigated, in particular, the restoration of dependencies and distributions density (pattern recognition) and the interpretation of the indirect experiments results. As already noted, the speech signal parameters are inherent in drift, which over time leads to a decrease in the qualitative characteristics of the ASRSCU. In studies [4.5] a hypothesis is formulated on the insignificant effect of the drift of the speech signal parameters on the quality of speaker recognition and is proposed to be compensated by introducing a number of corrective coefficients. So in [6] on the basis of the assumption of constant in time, but a small absolute value of the drift of the speech signal characteristic parameters, its probabilistic estimation is based on the study of the sequence of recognition sessions results and the upper limit of the degree of drift with the given error probability is estimated and taking into account the recognition system classifier training algorithm, but the question of the influence of the number of evaluated data on the reliability of the estimates isn't investigated. The paper [7] describes a method for determining the maximum drift rate allowed for a corresponding recognition system classifier, among which, however, there are no neural networks. In papers [8, 9], the phenomenon of drift recognizes and formulates the requirements for optimizing the parameters of the speaker recognition system classifier re-training process stating, in particular, the requirements regarding the phonetic composition of language materials for retraining, thereby reducing the total amount of study sample. In work [10] the influence of "extreme" variations of speech signals on the quality of the speaker recognition system is estimated, and the permissible limits of variation of spectral individual parameters are estimated. However, in all the aforementioned works, a priori assumptions are made about the nature and parameters of drift in speech signals, therefore, an urgent task is the generalization of the theory of statistical learning to the problem of a speech signal parameters drift in the long-term operation of the speaker recognition system.
3 MATERIALS AND METHODS
In an unknown distribution P(x, y) you can only
l
estimate the empirical risk Re = l-1 ^ Q(zi). This goal
i=1
can be achieved by adjusting the parameters of the
© Bisikalo O. V., Kovtun V. V., Yukhimchuk M. S., Voytyuk I. F., 2018 DOI 10.15588/1607-3274-2018-4-7
classifier a according to the condition a* = argminRe(a,l), evaluating the complexity of the
a
training process as G(l) , 2 ' 'f l <d, , where
[< ld (d!)-1, ifl > d,
d +1 - the smallest size of the set, under which the
condition G(l) = 2l is violated.
On the basis of the foregoing, we consider the problem of minimizing the risk functional
R(a)=JQ(z,a)P(z), aeA, (1)
as a task of minimizing the functional of empirical risk
l
Re(a) = lQ(z,,a), aeA (2)
i=1
over a set of indicator functions Q(z, a) = {0,1}. In this case (1) is characterized by the probability of incorrect classification Aa = {Q(z, a) = 1}, and (2) - by the frequency of occurrence of such an event. If all empirical data z are taken from the same distribution, then with probability 1 - n simultaneously for all N functions from a set Q(z, ak), k = 1,2,...,N, inequality is performed
R(ak )< Rmax(ak ) = Re (ak )+1"1(lnN - lnp)"
(i+ V1 + 2Re (ak )(lnN - ln p)-1 j. (3)
The equation (3) allows us to describe the dependence of the incorrect classification risk on the factor space dimension, which can be reduced by applying, in particular, the principal component analysis [11], thus reducing the computational complexity of the recognition task. However, in the context of the critical use of the recognition system, the increase in the wrong classification risk is unacceptable, which can be prevented by removing examples that increase risk from the training sample. This operation is proposed to be carried out on the basis of Shannon's informativeness [12]:
Rg = R(c) + k1 (d(ln(2d-1 (p - c))+ 1)s(ln(pc-1)+1 ))•
(P - c )-1.
If the parameters of the initiating probability distribution are unknown, then it is suggested to use the test to identify the distribution point in the context of the speaker's identity, which will be recognized by the ASRSCU [13]:
+(P1,Pi)= sup(P -P2I(min(0.5(P1 + P20.5(( + P2)))-05), (5)
A
where sup|S1 - S^ = 0.5dict((1, S2) - the empirical measure of the distance between the samples S1 and S2 . This test also allows us to determine the marginal sample size Pl ((S, P)>e)<(2l)de■0■25lsl and the boundary of the second kind of errors probability
P 2l (( S 2 )>8)<(2l)de■0■25l£l. (6)
In addition to the parameters of the training procedure on the quality parameters of the ASRSCU, the drift of the speech signal parameters is also influenced by the physiological changes in the speech apparatus of the human. If ASRSCU will operate for a long time, the quality of recognition will decrease, as the initiating distribution of input data will change. Next, we call this phenomenon a drift of a compatible distribution P(x, y). The theory of machine learning regulates the definition of the adequacy of the amount of training sample O for
"drifting" data in the form O = s-2(d + log8-1) [13]. However, the relevance of the relationship of drift with the empirical and true risk is relevant. Assume that the recognition is performed by the Bayesian classifier in terms of deterministic connection P(x, y )e{0,1). If the example provided for classification x, is close to the the training sample data min||x + x,-1| < s, then the
classification is carried out in accordance with a reliable estimate of conditional probability, and the error probability is min{p(x, A), P(x, B)} 0, and informative
xi - 1 (k)~ ^ 1 (k)=- log Pk (yk = ykW z2,•••, zk-b xk )
which defines the deviation degree of the input data from the data of the training sample. That is, if x, is mach different from the data of the training sample, then the error probability can be estimated as 2P(A)P(B)« 0.5, and informative I(k)« 2. In the field of drift, these indicators will generally take a form max{P(x, A), P(x, B)} 1, I (k . Consequently, there is a link between the need for re-training of the classifier and the value of empirical risk, embodied in the value of informativity. When creating critical systems, risk management is necessary, so we will combine the retraining operation with the situation of exceeding the value of the empirical risk of some threshold. That is, we will carry out repeated training if with a probability p in
at least one of an N functions Q(z,ak), k = 1,—,N the upper limit of risk exceeds the thresholds by the testing results on the sample no less than from the m elements: Rmax (a k )> Rm, or by revealing this relation:
Re (ak) + K + V1 + 2Re (ak )K- j > Rm , where
K¡ = (ln N - ln p)-1, l > m. Given the previous transformations we obtain Re (ak) > Rm 2RmK¡ , or
Re (a k )> Rm -V 2Rm (ln N - ln p)-1
(7)
Inequality (7) describes the ratio of empirical risk to threshold values for any one l > m . If the recognition system is used, then based on the generalization of the results of its work for a certain time you can calculate the empirical risks Re (a k) for the various classes of system parameters, for example, the length or content of the
passphrase, the number of microphones for its recording, acoustic space parameters where the system is operating, etc. If for some class the empirical risk exceeds the threshold Re (ak) > Rtr, then for the data of this class, you need re-training.
Let's describe the probability of a situation when after testing a system on a set of elements the empirical risk will exceed the threshold Rtr :
i „. , , \
(8)
= sup
1<k <N
X
I=\R,r
^\Rak )Rtrl](1 - R(ak ))Mtr )J
where unknown true risks R(ak) are used, the limit values of which can be obtained by analyzing the empirical risks, but taking into account the limited sample size, these estimates will be understated. However, it is possible to obtain a reliable lower boundary r providing for a monotonous increase in the likelihood of re-training Pr with growth R(ak). Let ((n, n > 1) be a sequence of intervals between the re-training procedures of the recognition system, measured in the number of recognition sessions performed. Assume that -independent randomly distributed probability variables, then the probability that re-training will occur through
sessions will describe by p{n = t} = qt-1Pr, where q = 1 - Pr, and the probability that re-training will occur no more than through the sessions we describe like p{n < t} = 1 - q . The mathematical expectation of the interval between sequences of re-training will describe by
W W f W A
T = ^(iin )=ztqt-1Pr =E Z q
t=1 t=1v*=t
J-1
x-1
Pr =z(-1)1 - q)-1 Pr
t=1
= X qt 1 = Pr 1, from where the expected
average
t=1
operation duration of the recognition system without retraining is lm = mx = mPr-1. Also, we obtain the mathematical expectation of the number of re-training
W
procedures for sessions: H(()=Z p(n < t),
«=1
lim H (t )t_1 ='
= Pr
Determine the effect of re-training on some limited positive rational function G > g(() > 0 taking into account the corrective operator T. For a given Pr correction, the evaluation (8) is based on the decomposition of a complex phenomenon on a complete set of incompatible events p(a) = z p( )p(a|B ):
B
p| sup R(a k)- Re (a k )((a k ))^0'5 >e| Li<kJ . (9)
< NPr a(1 - a)-1 + N(1 - Pr (1 - a)-1 ))t,
where a = qe 0 5e m . Using inequality (9) we establish a connection between e and reliability taking into account that
p0 > NPra(1 - a) 1 ^ a < p0(p0 + NPr ) 1 ^ qe
-0.5e2m
:P0 (0 + NPr)
-1
e > en = -v - 2
V- 2m-1ln( (( + NPr ) - Pr ))-1 ). (10)
We simplify inequality (9) taking into account the fact that the values a and (1 - a)-1 decreases with growth e determine what value ensures reliability 1 - p, p > p0, taking into account that
p > P0 + Nqa ^ q e
t -0.5e2m'
(p-u0 )(Nq) 1 ^-0.5e2m
< ln((p-p0 )(n-1q"(t+1))):
e > e1 = J- 2m- [ln[((p - p0 )N-)- ((+ 1)ln(1 - Pr)). (11)
Using the above considerations, we obtain on the basis of (9) taking into account (11) the expression to determine the true risk:
R(a k )< Re (a k ) + e^1 + ^1 + 2Re (a k )e11 I =
= Re (ak ) + V- 2m"1 ln((p - p0 )n _1q"((+1)) x (12) 1 + 1 + 2Re (ak)(- 2m-1 ln((p - p0 )N -1q-(i+1)))°.
and generalizing (10) and (11) we obtain the constraints on the choice p0 for (12):
0.5N( + q')+ 0.5pf^/4NPr +((1((. + q')))) . (13)
where e(l) = 4hl- (ln(2lh-)+1)- 4l- ln(0.24p).
If the re-training procedure is implemented, it's expedient to minimize the true risk, and estimate the limit of the empirical risk for the specified re-training risk. It
has been previously grounded that re-training of the classifier with reliability 1 - p willn't occur, if q' > 1 - p , that allows (14) to obtain an analytical expression for calculating the boundary that describes the effect of retraining on the choice of a classifier:
Rr < Re + 0.5e(m)^1 + ^ 1 + 4Re (e(m)) 1
■^0.5e(m _ln(1 - p)(ln(1 - Pr )) 1 ).
(15)
p>p0 >1
We formulate measures on the practical use of the aforementioned theory of the risk assessment of ASRSCU taking into account the procedure of the classifier retraining as a result of the drift of the speech signal parameters. In the context of the foregoing, ASRSCU requires a classifier designed to take into account the balance between the reduction of the empirical risk and the increase in the difference between the empirical and true risk with increasing complexity of the classifier, by which we mean the capacity of the set of input data that the classifier is capable of recognizing. The indicated balance is proposed to be ensured by minimizing the upper limit of true risk for the specified values of reliability and duration of the training sample. Based on [14] we formulate a kind of indicator function that minimizes empirical risk with probability 1 - p :
Re -V0.5e(() < R < Re + 0.5e(()(l ^ 1 + 4Re(e(l))-1 ] , (14)
Consequently, the authors proposed a set of measures for assessing the operational risks of long-term use of ASRSCU. In particular, using (3) describes the dependence of the risk of incorrect classification from the dimension of the factor space. Based on the formulated measure of informativity with the help of (4) it is possible to analyze the study sample on the presence of examples that lead to increased risk. With the help of (8) we describe the influence of the phenomenon of drift of input parameters on the qualitative performance indicators of the ASRSCU, and with the help of (13) an estimation of the operation duration of the ASRSCU is performed, during which it is impractical to re-train the classifier. Also (15), it is possible to choose the optimal classifier on the position of minimizing its complexity, taking into account the risks of long-term use of the ASRSCU and the possibilities of re-training. In general, the above-mentioned material for the first time comprehensively describes the problem of minimizing the average operation risk of the ASRSCU under empirical data, generalized taking into account nonstationary input data with drift patterns and parameters of the recognition system classifier. The limits of confidence intervals of risk are calculated taking into account the procedures of classifier re-training.
4 EXPERIMENTS
The statistical data for the empirical assessment of the adequacy of the above theoretical concepts for the operational risks analysis of the ASRSCU is obtained on the basis of the analysis of the results of long-term use of ASRSCU at the Department of Computer Control Systems of Vinnytsia National Technical University. The mentioned ASRSCU has a classical architecture, which includes a block of preliminary speech signal processing, a block of informative features allocation and a classification block.
In the pre-processing block, the detection of speech activity intervals in phonograms was performed using a two-channel VAD algorithm [16]. Intervals of linguistic activity lasting 3 seconds were segmented into frames of duration 30 ms with 15 ms shift. To compensate for the Gibbs effect, the signal was weighed by the Hemming window. Effects of channel distortions at the factor level were offset by the calculation of the cepstral mean subtraction and, taking into account the sufficient duration of the analysis frameworks, the implementation of the feature warping [17].
+
+
x
In the block of informative features extraction from each of the received from the block of preliminary processing frames extracted 19 normalized by the power cepstral coefficients [18], their energy and their first and second derivatives. Also, for the presentation of speech signals, the position of the theory of spectral-temporal receptive fields was used, which describes the work of the human auditory system with the involvement of the results of psychoacoustic and neuropsychological studies of the peripheral and central auditory system of mammals in the spectral and temporal spaces [19, 20]. The STRF-description of the speech signal included two stages. At the first stage, the auditory spectrum was obtained as a result of the simulation of the peripheral auditory system. At the second stage on the basis of the first stage results the high-level representations of linguistic representations as the results of simulation of the auditory cortex of the central nervous system of human ware synthesized.
For the implementation of the first stage, an affine wavelet transform yc ((, f) of the speech signal frame s((), was initially carried out, which was passed through a bank of cochlear filters: yc (', f) = s(()*t h((, f), where
*' - is a convolution operation in the time space. Further, the work of hair cells yA (t, f) was modeled, which was consistently performed: the operation of high-frequency filtration to emulate the process of converting sound pressure into the speed of hairs; nonlinear compression operation g(u); low frequency filtering operation w(t) to emulate phase blocking of the auditory nerve: yA (', f ) = g(dtyc (', f ))*'w('). Next, the work of the lateral inhibitory network of the cochlear nucleus yLIN (t, f) was modeled in the form of a frequency selection operation, for which the partial derivative of the yA (t, f) frequency was passed through a half-period rectifier: yLIN (t, f ) = max(df-yA ((, f ),0). And the first
stage was ended by receiving the auditory spectrogram y(t, f) by convolution yLiv ((, f) in the time space with a short-term window function |(t, k 2):
y(t,f) = yLIN(t,f)*t l(t,K2), where |((,t)= e"tK2 '. In Fig. 1 we can visually compare examples of Fourier and STRF-spectrograms.
4 it*
500
1000 1600 2000 TitriB (ms)
500 1000 1500 3000 2500
Tims (ms)
500
250
125
1000 1500 2000 Time (ma)
1000 1500 2000
Time (ms)
Figure 1 - Visual representation of a speech signal: a, b - Fourier spectrographs of the speech signal without / with noise respectively; c, d - STRF spectrograms of the speech signal without / with noise respectively
The ongoing second phase was based on obtaining the STRF-function as a result of the combination hs and hT :
strf = hs • ht, where hS(/,o,e)= hsc(,ro)cose+hcc(f,o)sine,
hT ((, Q, 9) = hrt ((, Q)cos 9 + hrt (t, Q)sin 9 . Operation
STRF((, f, Q, o, 9, e) on a spectrograph y(t, f) describes
like STRF((, f,Q,o,9,e) = y(t, f ) [hs(,o,e))(t,Q,9)],
where *f is the convolution operation in both time and
frequency spaces. Fig. 2 shows a scalable STRF representation of one of the speech signal frames from Fig. 1 in hsc /hrt space and MFCC- representation of the same frame.
In the investigated ASRSCU from the frames of the speech signal, according to the results of the STRF analysis, three informative features were distinguished. The first feature F1((, o) was obtained by summing the values of all elements of the STRF representation in hsc/hrt spaces:
F1 ((, o)=ZZ I STRF (t, f, Q, 0,0,0) ,
f Q
where o = 1,2,..., No , No — is the number hsc elements, and the values of the phase parameters 9 and e , given their small informativity for a speaker recognition task
(16)
[19], was considered equal to 0 for a simplification of calculations. The second feature was obtained by logarithm F1(t, o):
F2 (t, o) = log(F1((, o)). (17)
The third STRF feature was obtained using the discrete cosine transform (DCT) [20] to F2 ((, o):
F3 ((, k)= Z F2 ((, o)cos(2nokN(-1), (18)
o=1
where k = 1,2,.,Nk , Nk < No .
Thus, the vector of informational attributes for one frame of the input speech signal after its processing consisted of 79 elements that are visually represented in the form of a spectrogram-like structure, where the axis of the ordinates is postponed by the number of frames along the abscissa, the values of the ordinate axes correspond to the numbers of informative features, and the intensity of the color shows the value of the corresponding features within the frame, multiplied by the corresponding weighting factor. Such a way of presenting informative features is due to the type of ASRSCU classifier.
hn 1 00
hn 1 00
-32.-16.-8.-4 ■ 2. -
2. 4 8 IS 32
Figure 2 — Visual representation of a speech signal: a, b — MFCC-presentation of speech recording without noise / with noise; c, d -STRF-representation of spectrographs of speech recording without noise / with noise in space hsc jhrt
In the classification block of the ASRSCU, a convolutional neural network [21] was implemented. Its architecture is designed taking into account the recommendation (15) regarding the complexity of the recognition system classifier, taking into account the features space parameters, operating conditions and the purpose of the ASRSCU. The structure of network (see Fig. 3) includes two convolution layers for features extraction, two sub-sampling layers to reduce features dimension, two local normalization layers, three full-connected layers and finalized by an output SOFTMAX layer.
The convolutional layer of the neural network performs a two-dimensional convolution operation of the fragment of the input image and a filter to extract the height-level features based on the activation functions of the Rectified Linear Units:
of the central pixel of the input fragment xm to C
(
yn = max
m"
\
0, I
wU * xU + bU
To reduce the data
dimension, the sub-sampling MAX-pooling type layers are used, at the outputs of which the maximum values of square pieces are obtained in size 3 x 3, on which the input card is broken off without overlapping. In order to prevent a decrease in the network training process speed the Local Response Normalization procedure
implemented, in which the normalized response BU (x, y) at the output of yU (x, y) MAX-pooling sub-sampling layer at the position (x, y) is obtained as BU (x, y) =
= H" (x, y)
min (m" ,n+0.5k 4 ) ^ K6
K3 +K5 I H" (x, y)2
n=max(0,n-0.5 k 4 )
where Mu — is a total number of cores in the layer u ,
-3
and the values k3 = 2 , k4 = 5 , k5 = 10 , k6 = 0.75 defined empirically. Dropout technology is implemented to prevent overhaul on the full-connected network layer. SOFTMAX classifier on the output layer of the network determines the probability distribution ym of membership
speakers classes like ym = e
I eX "
V n=1
where
bn ), M = 100 . For network training
M
"Kn = Z (wn,m * xm +
m=1
a stochastic gradient descent method with step 128 was used. The rule for updating weight wk on k iteration
looks like Wk+1 = Ak+1 + Wk, where
Ak+1 = 0.9Ak - 0.004k7Wk -K7 dZ/dwk and S^/dwk is a derivative. The initial values of the neuron weights on each layer were set using the zero mean Gaussian distribution with a standard deviation of 1. The training error was 0.0002.
5 RESULTS
The main purpose of the experiments carried out with the above-described ASRSCU was to assess the impact of the operation duration on the recognition system quality performance with the generalization of data on the informativity of the attributes space elements. For this purpose, the ASRSCU software was installed on three computers at the Department of Computer Control Systems of the Vinnytsia National Technical University, which operated for two years. Experiments were attended by 6 speakers (4 male and 2 female), each of whom conducted regular recognition sessions at least once every five days (total of over 2000 recognition sessions per speaker per study period) with fixation of results . The possible result of the recognition session was the correct speaker's recognition, the speaker's confusion (the first kind error, Miss) or denial access (second kind error, False Alarm). The results of experiments are presented in the form of detection error trade-off curves, which show the dependence of the likelihood of the first kind errors Pa occurrence from the second kind errors PR occurrence
CI:
24 maps *15*15
Figure 3 — Architecture of the ASRSCU's convolutional neural network classifier © Bisikalo O. V., Kovtun V. V., Yukhimchuk M. S., Voytyuk I. F., 2018 DOI 10.15588/1607-3274-2018-4-7
1
m
probability, with the same threshold decision making recognition system's classifier. In particular, Fig. 4 shows the DET curves depending Pa/ Pp on the operation duration of
the ASRSCU without the re-training of the classifier, to evaluate the drift of the individual features that characterized the speakers in the recognition process.
Fig. 5 shows the DET curves for Pa/ Pp compliance
with the recommendations for the re-training frequency
HO
40
20
10
2 1
0.5
0.2 0.1
0.001 0.01 0.10.2 0.5 1 2 5 10 20 40
Pp,%
Figure 4 - DET curves PajPp of ASRSCU depending on the duration of operation without re-training of the classifier
P„,% 40
20
10
2
1
0.5
02 0.1
- -r£-trailUlli;-iiianLre$ drift
---TTe*training-f<atiire$ drift
---w-tiniisina+fe^tiLres drift
0.01 0.10 2 0.5 1 2 5 10 20 40
Pp,%
Figure 5 - DET curves PajPp of the ASRSCU, depending on
the observance of the recommendations for the re-training frequency and the length of the training sample, taking into account the drift of individual speech parameters
and the length of the training sample, taking into account the drift of individual speech parameters. Parameters of the periodicity of re-training and the length of the training sample were determined for the ASRSCU by the formulas (8) and (13) respectively.
Fig. 6 shows the DET curves Pa/ Pp depending on
the configuration of the ASRSCU features space, which regularly passed re-training procedures with the parameters of the training sample, regulated by the above theoretical results.
p„,% 40
20
10 5
2 I
0.5
0.2 0.1
- strf ftlturti
---: ( feflruits
- - STRf^fNCC fMBmt
0.001 0.01 0.10.2 O S 1 2 5 10 20 40
Pp.%
Figure 6 - DET curves PajPp of the ASRSCU, depending on the feature space configuration
The obtained DET curves confirm the theoretical assessments adequacy of the sufficiency of the classifier's complexity to make decisions on the speaker personality of the ASRSCU, confirming the expediency of the retraining procedure of the ASRSCU classifier and correctness determined on the basis of theoretical estimates of this procedure parameters.
6 DISCUSSION
The results of the experiments on a Fig. 4 show that the quality indicators of the ASRSCU during a long-term exploitation process are reduced stochastically without the possibility of identifying an adequate tendency that to some extent allows the use of the speaker recognition system for its intended purpose, but makes its critical application impossible, one of the conditions of which is predictability of the work results.
The results shown in Fig. 5 clearly confirm the expediency of the re-training procedure of the classifier, whose parameters are regulated by the theoretical results obtained in part 4 of the article. It should be noted that, in addition to observing the periodicity of re-training, the obtained results reveal the relationship between the ASRSCU's first and second kind errors probabilities and the composition and the size of the training samples used
for re-training. On the basis of the results analysis of a long-term exploitation of the ASRSCU, the effect the informative features drift on the quality of the system's operation was found which provides objective material for optimization of the ASRSCU factor space by reevaluating the weight of the informative features, which are subsequently visualized before using the convolutional neural network classifier.
The results shown in Fig.6 on the one hand show a greater informativeness of the features based on the power-normalized cepstral speech signals analysis. However, the features that result from the practical application of the theory of spectral-temporal receptive fields make up only about 4% of the features space, but not only can significantly increase the quality of the ASRSCU, but also make the DET curves more linear, that is, in general, stabilize the decision-making process by the system critical use.
CONCLUSIONS
In the article a theoretical analysis of the long-term operation process of the ASRSCU was conducted, on the basis of which the practical recommendations for the stabilization of the quality indicators of the recognition system are formulated.
The scientific novelty of the obtained results can be attributed to the fact that for the first time a theoretical analysis of the problem of an average risk minimization has been made on the empirical operation results of the speaker recognition system for critical use, in which, unlike the existing approaches, non-stationary input data with drift patterns and characteristic features of the recognition system classifier are taken into account, which allowed to estimate the limits of the risk confidence intervals, provided that the re-training sessions were carried out. The practical consequence of the theoretical analysis is the formulated set of measures for assessing the operational risks of long-term use of the ASRSCU. In particular, using (3) the dependence of the wrong classification risk on the dimension of the factor space is described. Based on the formulated measure of informativity (4), an analysis of the training sample on the identification of elements that lead to increased risk was made. Using (8), we describe the influence of the phenomenon of drift of the speech signals parameters on the qualitative performance indicators of the ASRSCU. With the help of (13), an estimation of the operation duration of the ASRSCU was carried out, during which it was impractical to re-train the classifier. Applying (15) the optimal classifier was chosen from the position of minimization of its complexity, taking into account the risks of long-term use of ASRSCU and the possibility of re-training. In particular, the resulting ASRSCU-based convolutional neural network classifier has a compact structure and confirmed the predicted efficiency. The formulated recommendations correctness is confirmed by empirical results presented in the form of DET curves.
Subsequent studies are planned to devote to the detection of the final potential of the spectral-temporal
receptive fields theory in the context of the informative features for speaker recognition synthesis. As the results of experiments have shown, their use not only significantly increases the qualitative performance of the ASRSCU, but also make the DET curves more linear, that's, in general stabilizes the decision-making process by a system of critical use. It is planned to investigate the potential of introducing into the list of information features used in the ASRSCU the human speech source parameters and to make the final factor space optimization.
ACKNOWLEDGEMENTS
The work was carried out within the framework of the cathedral scientific research work number 46K4 "Methods of modeling and optimizing complex systems on the basis of intellectual technologies" at the Department of Computer Control Systems of the Vinnytsia National Technical University with the support of the department staff and the staff of related Department of Automation and Information Measuring Technologies.
REFERENCES
1. Kovtun V. V., M. M. Bykov Ocinjuvannja nadijnosti avtomatyzovanyh system rozpiznavannja movciv krytychnogo zastosuvannja, Visnyk Vinnyc'kogopolitehnichnogo instytutu, 2017, No. 2, pp. 70-76.
2. Speaker verification over the telephone [Electronic resource]. Access mode: https://pdfs.semanticscholar.org/cad0/bfdec3f4fb1198f63c959580d7 217d541a0f.pdf
3. Introduction to Statistical Learning Theory [Electronic resource]. Access mode: http://www.kyb.mpg.de/fileadmin/user_upload/files/publications/pd fs/pdf2819.pdf
4. Learning deep architectures for AI [Electronic resource]. - Access mode: https://www.iro.umontreal.ca/~bengioy/papers/ftml_book.pdf
5. Scaling learning algorithms towards AI [Electronic resource]. Access mode: http://yann.lecun.com/exdb/publis/pdf/bengio-lecun-07.pdf
6. Learning a similarity metric discriminatively, with application to face verification [Electronic resource]. Access mode: http://yann.lecun.com/exdb/publis/pdf/chopra-05.pdf
7. Jang G., Lee T., Oh Y. Learning statistically efficient feature for speaker recognition, IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 7—11 May 2001: proceedings. Salt Lake City, UT, USA: IEEE, 2002, pp. 4117-4120. DOI: 10.1109/ICASSP.2001.940861.
8. Unsupervised feature learning for audio classification using convolutional deep belief networks [Electronic resource]. Access mode: http://www.robotics.stanford.edu/~ang/papers/nips09-AudioConvolutionalDBN.pdf
9. Learning methods for generic object recognition with invariance to pose and lighting [Electronic resource]. Access mode: http://yann.lecun.com/exdb/publis/pdf/lecun-04.pdf
10. Learning a nonlinear embedding by preserving class neighbourhood structure [Electronic resource]. Access mode: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.124.8635 &rep=rep1&type=pdf
11. A tutorial on Principal Components Analysis [Electronic resource]. Access mode: http://www.cs.otago.ac.nz/cosc453/student_tutorials/principal_com ponents.pdf
12. Gustafson J. L., Montry G. R., Benner R. E. Development of parallel methods for a 1024-processor hypercube, SIAM Journal on Scientific and Statistical Computing, 1988, Vol. 9, No. 4, pp. 609-638.
13. Bell J., Casasent D., Bell C. G. An Investigation of Alternative Cache Organizations, IEEE Transactions on Computers, 1974, Vol. C-23, No. 4, pp. 346-351.
14. Sergienko I. V. Topical directions of informatics. In memory of V. M. Glushkov. New York, Heidelberg, Dordrecht, London: Springer, 2014, 286 p.
15. Sampling - 50 years after Shannon [Electronic resource]. Access mode: http://bigwww.epfl.ch/publications/unser0001.pdf
16. Mak M. W., Yu H. B. A study of voice activity detection techniques for NIST speaker recognition evaluations, Computer, Speech and Language, 2014, Vol. 28, No. 1, pp. 295-313. DOI: 10.1016/j.csl.2013.07.003.
17. Front-end factor analysis for speaker verification [Electronic resource]. Access mode: http://habla.dc.uba.ar/gravano/ith-2014/presentaciones/Dehak_et_al_2010.pdf
18. Power-normalized cepstral coefficients (PNCC) for robust speech recognitions [Electronic resource]. Access mode: http://www.cs.cmu.edu/~robust/Papers/OnlinePNCC_V25.pdf
УДК 681.327.12
19. Speech Processing with a Cortical Representation of Audio [Electronic resource]. Access mode: https://pdfs.semanticscholar.org/f1d8/f93cdb64390b3a65f930cee43 46c30bd86e4.pdf
20. Using spectro-temporal features to improve AFE feature extraction for automatic speech recognition [Electronic resource]. Access mode:
https://pdfs.semanticscholar.org/c7c5/04087f2107f0ea9a3cedeeaf5c c0c48c0c92.pdf
21. Kovtun V. V., Bykov M. M. Doslidzhennja efektyvnosti oznak rozpiznavannja movciv pry vykorystanni zgortal'nyh nejromerezh,
Optyko-elektronni informacijno-energetychni tehnologii', 2016, No. 2(32), pp. 22-28.
Received 03.06.2018. Accepted 25.07.2018.
АНАЛ1З РЕЗУЛЬТАТ1В ЕКСПЛУАТАЦП ВТОМАТИЗОВАНО1 СИСТЕМИ РОЗШЗНАВАННЯ МОВЦЯ КРИТИЧНОГО
ЗАСТОСУВАННЯ
Бкчка.ю О. В. - д-р техн. наук, професор, декан факультету комп'ютерних систем i автоматики Вшницького нацюнального техшчного ушверситету, Вшниця, Укра!на.
Ковтун В. В. - канд. техн. наук, доцент, доцент кафедри комп'ютерних систем управлшня Вшницького нацюнального техшчного ушверситету, Вшниця, Укра1на.
Юхимчук М. С. - канд. техн. наук, доцент, доцент кафедри комп'ютерних систем управлшня Вшницького нацюнального техшчного ушверситету, Вшниця, Укра1на.
Войтюк I. Ф. - канд. техн. наук, доцент кафедри комп'ютерних наук Терношльського нацюнального економiчного ушверситету, Терношль, Укра1на.
АНОТАЦ1Я
Актуальтсть. У статт узагальнюеться теорiя статистичного навчання для оцшювання результата довготривало! експлуатаци автоматизовано! системи розпiзнавання мовця критичного застосування (АСРМКЗ) iз урахуванням особливостей об'екту, iз яким працюе система, та структурно! специфжи такого класу систем розпiзнавання.
Мета роботи. Розроблення цiлiсного комплексу заходiв для стабшзацп якiсних параметрiв АСРМКЗ при !! довготривалш експлуатаци.
Метод. У робот! сформульовано комплекс заходiв для оцшювання експлуатацшних ризикiв тривалого використання АСРМКЗ. Зокрема, описано залежшсть ризику неправильно! класифшаци вiд розмiрностi факторного простору. Базуючись на сформульованiй мiрi шформативнос™, проаналiзовано заходи щодо аналiзу навчально! вибiрки для виявлення прикладiв, як призводять до зростання ризику. Анаттично описано вплив явища дрейфу параметрiв мовних сигналiв на якiснi показники ефективносп АСРМКЗ. Здiйснено оцiнювання тривалост експлуатацi!' АСРМКЗ, на протязi яко! здiйснювати повторне навчання класифiкатора недоцшьно. Сформульовано рекомендацi!' щодо вибору оптимального класифжатора АСРМКЗ з позицi! мiнiмiзацi! його складност iз урахуванням ризиюв тривало! експлуатацi! АСРМКЗ та можливктю процедури повторного навчання.
Результати. Шдтверджено адекватнiсть отриманих у роботi теоретичних результатов представленими у виглядi БЕТ-кривих даними, якi узагальнюють iнформацiю вщ довготривалих експерименпв iз АСРМКЗ, у якш при формуваннi конфiгурацi! простору ознак враховувалися нормованi за потужшстю кепстральнi коефiцiенти та похiднi вщ них характеристики i ознаки, отримаш на основi теорi! спектрально-темпоральних рецептивних полiв. В рамках створено! теоретично! концепцп проведено оцiнювання впливу конфЬураци простору ознак та виду i складностi класифiкатора на стабiльнiсть яккних параметрiв АСРМКЗ при !! довготривалш експлуатацi!.
Висновки. Вперше теоретично проанал!зовано проблему мiнiмiзацi! середнього ризику по емшричним результатам експлуатацi! системи розшзнавання мовця критичного застосування, де, на вщмшу вiд iснуючих пiдходiв, враховано нестацюнаршсть вхiдних даних iз дрейфом ¡ндивщуальних параметрiв мовних сигна^iв та характеристичш параметри класифiкатора системи розшзнавання, що дозволило оцiнити межi довiрчих iнтерва^iв ризику за умови здшснення сеансiв повторного навчання.
КЛЮЧОВ1 СЛОВА: автоматизована система розшзнавання мовщв критичного застосування, планування експерименту, факторний аналiз, теорiя статистичного навчання.
УДК 681.327.12
АНАЛИЗ РЕЗУЛЬТАТОВ ЭКСПЛУАТАЦИИ АВТОМАТИЗОВАННОЙ СИСТЕМЫ РАСПОЗНАВАНИЯ ДИКТОРА
КРИТИЧЕСКОГО ПРИМЕНЕНИЯ
Бисикало О. В. - д-р техн. наук, профессор, декан факультета компьютерных систем и автоматики Винницкого национального технического университета, Винница, Украина.
Ковтун В. В. - канд. техн. наук, доцент, доцент кафедры компьютерных систем управления Винницкого национального технического университета, Винница, Украина.
Юхимчук М. С. - канд. техн. наук, доцент, доцент кафедры компьютерных систем управления Винницкого национального технического университета, Винница, Украина.
Войтюк И. Ф. - канд. техн. наук, доцент кафедры компьютерных наук Тернопольского национального экономического университета, Тернополь, Украина.
АННОТАЦИЯ
Актуальность. В статье обобщается теория статистического обучения для оценки результатов длительной эксплуатации автоматизированной системы распознавания диктора критического применения (АСРДКП) с учетом особенностей объекта, с которым работает система, и структурной специфики такого класса систем распознавания.
Цель работы. Разработка целостного комплекса мер по стабилизации качественных параметров АСРДКП при ее длительной эксплуатации.
Метод. В работе сформулирован комплекс мер для оценки эксплуатационных рисков длительного использования АСРДКП. В частности, описано зависимость риска неправильной классификации от размерности факторного пространства. Основываясь на сформулированной степени информативности, сформулированы меры по анализу обучающей выборки для выявления примеров, которые приводят к росту риска. Аналитически описано влияние явления дрейфа параметров речевых сигналов на качественные показатели эффективности АСРДКП. Осуществлена оценка зависимости длительности эксплуатации АСРДКП, в течение которой осуществлять повторное обучение классификатора нецелесообразно. Сформулированы рекомендации по выбору оптимального классификатора АСРДКП с позиции минимизации его сложности с учетом рисков длительной эксплуатации АСРДКП и возможности процедуры повторного обучения.
Результаты. Подтверждена адекватность полученных в работе теоретических результатов представленными в виде БЕТ-кривых данными, которые обобщают информацию от длительных экспериментов с АСРДКП, в которой при формировании конфигурации пространства признаков учитывались нормированные по мощности кепстральные коэффициенты и производные от них характеристики и признаки, полученные на основе теории спектрально-темпоральных рецептивных полей. В рамках созданной теоретической концепции проведена оценка воздействия конфигурации пространства признаков и вида и сложности классификатора на стабильность качественных параметров АСРДКП при ее длительной эксплуатации.
Выводы. Впервые теоретически проанализирована проблема минимизации среднего риска по эмпирическим результатам эксплуатации системы распознавания диктора критического применения, где, в отличие от существующих подходов, учтена нестационарность входных данных с дрейфом индивидуальных параметров речевых сигналов и характерные параметры классификатора системы распознавания, что позволило оценить пределы доверительных интервалов риска с условием осуществления сеансов повторного обучения.
КЛЮЧЕВЫЕ СЛОВА: автоматизированная система распознавания диктора критического применения, планирование эксперимента, факторный анализ, теория статистического обучения.
Л1ТЕРАТУРА / ЛИТЕРАТУРА
1. Ковтун В. В. Ощнювання надшност автоматизованих систем розшзнавання мовщв критичного застосування / М. М. Биков, В. В. Ковтун // Вкник Вшницького полгтехшчного iнституту. -2017. - № 2. - С. 70-76.
2. Speaker verification over the telephone [Electronic resource]. -Access mode: https://pdfs.semanticscholar.org/cad0/bfdec3f4fb1198f63c959580d7 217d541a0f.pdf
3. Introduction to Statistical Learning Theory [Electronic resource]. -Access mode: http://www.kyb.mpg.de/fileadmin/user_upload/files/publications/pd fs/pdf2819.pdf
4. Learning deep architectures for AI [Electronic resource]. - Access mode: https://www.iro.umontreal.ca/~bengioy/papers/ftml_book.pdf
5. Scaling learning algorithms towards AI [Electronic resource]. -Access mode: http://yann.lecun.com/exdb/publis/pdf/bengio-lecun-07.pdf
6. Learning a similarity metric discriminatively, with application to face verification [Electronic resource]. - Access mode: http://yann.lecun.com/exdb/publis/pdf/chopra-05.pdf
7. Jang G. Learning statistically efficient feature for speaker recognition / G. Jang, T. Lee, Y. Oh // IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 7-11 May 2001: proceedings. - Salt Lake City, UT, USA: IEEE, 2002. - P. 4117-4120. DOI: 10.1109/ICASSP.2001.940861.
8. Unsupervised feature learning for audio classification using convolutional deep belief networks [Electronic resource]. - Access mode: http://www.robotics.stanford.edu/~ang/papers/nips09-AudioConvolutionalDBN.pdf
9. Learning methods for generic object recognition with invariance to pose and lighting [Electronic resource]. - Access mode: http://yann.lecun.com/exdb/publis/pdf/lecun-04.pdf
10. Learning a nonlinear embedding by preserving class neighbourhood structure [Electronic resource]. - Access mode: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.124.8635 &rep=rep1&type=pdf
11. A tutorial on Principal Components Analysis [Electronic resource]. - Access mode:
http://www.cs.otago.ac.nz/cosc453/student_tutorials/principal_com ponents.pdf
12. Gustafson J. L. Development of parallel methods for a 1024-processor hypercube / J. L. Gustafson, G. R. Montry, R. E. Benner // SIAM Journal on Scientific and Statistical Computing. - 1988. - Vol. 9, № 4. - P. 609-638.
13. Bell J. An Investigation of Alternative Cache Organizations / J. Bell, D. Casasent, C. G. Bell // IEEE Transactions on Computers. -1974. - Vol. C-23, № 4. - P. 346-351.
14. Sergienko I. V. Topical directions of informatics. In memory of V. M. Glushkov / I. V. Sergienko. - New York, Heidelberg, Dordrecht, London : Springer, 2014. - 286 p.
15. Sampling - 50 years after Shannon [Electronic resource]. - Access mode: http://bigwww.epfl.ch/publications/unser0001.pdf
16. Mak M. W. A study of voice activity detection techniques for NIST speaker recognition evaluations / M. W. Mak, H. B. Yu // Computer, Speech and Language. - 2014. - Vol. 28, № 1. - P. 295313. DOI: 10.1016/j.csl.2013.07.003.
17. Front-end factor analysis for speaker verification [Electronic resource]. - Access mode: http://habla.dc.uba.ar/gravano/ith-2014/presentaciones/Dehak_et_al_2010.pdf
18. Power-normalized cepstral coefficients (PNCC) for robust speech recognitions [Electronic resource]. - Access mode: http://www.cs.cmu.edu/~robust/Papers/OnlinePNCC_V25.pdf
19. Speech Processing with a Cortical Representation of Audio [Electronic resource]. - Access mode: https://pdfs.semanticscholar.org/f1d8/f93cdb64390b3a65f930cee43 46c30bd86e4.pdf
20. Using spectro-temporal features to improve AFE feature extraction for automatic speech recognition [Electronic resource]. - Access mode:
https://pdfs.semanticscholar.org/c7c5/04087f2107f0ea9a3cedeeaf5c c0c48c0c92.pdf
21. Ковтун В. В. Дослщження ефективносп ознак розшзнавання мовщв при використанш згортальних нейромереж / М. М. Биков, В. В. Ковтун // Оптико-електронш шформацшно-енергетичш технологи. - 2016. - № 2 (32). - С. 22-28.
