In this research, the task of extracting speech signal recognition features for voice identification of a person in a remote mode was solved, which imposes several restrictions, namely: (1) minimum processing time of the speech signal realization, since the required recognition reliability is achieved through statistical processing of the results; (2) reduction of the dimensionality of recognition features, since the process of extracting recognition features and their classification occurs on the transmitting side of the communication channel, which in turn imposes certain factors of computing power and noise in the communication channel. After analyzing the given conditions of the voice identification system, the question arose of developing a method for extracting speech signal recognition features that would provide more informative spectral characteristics of the speech signal, which would improve the efficiency of their further classification under the influence of noise. In this paper, we consider the possibility of applying the theory of time-scale analysis to solve this problem, namely, the development of a method for extracting recognition features based on the wavelet packet transform using the orthogonal basis wavelet function of Meyer and subsequent averaging of wavelet coefficients that are in the frequency band of the corresponding wavelet packet. Experimental studies have shown the ability of the developed method to generate speech signal recognition features with a close frequency-temporal structure based on wavelet packets in the Meyer basis, namely, it was found that at a signal-to-noise ratio of 10 dB, the features obtained based on the developed method have a very acceptable result, namely, 1.6-2 times more robust to noise than the features obtained based on the traditional Fourier spectrum, where the total deviation of the root mean square error of the obtained features is unacceptable at a signal-to-noise ratio of 20 dB.
The development of new methods and means of
ensuring information security is intended
primarily to prevent threats of access to
information resources by unauthorized persons.
To solve this problem, it is necessary to have
identifiers and create identification procedures
for all users. Modern identification and
authentication include various systems and
methods of biometric identification [
One of the most common biometric
characteristics of a person is his or her voice,
which has a set of individual characteristics that
are relatively easy to measure (for example, the
frequency spectrum of the speech signal). The
advantages of voice identification also include
ease of application and use, and the fairly low cost
of devices used for identification (e. g.,
microphones) [
Voice identification capabilities cover a very
wide range of tasks, which distinguishes them
from other biometric systems. First of all, voice
identification has been widely used for a long
time in various systems for differentiating access
to physical objects and information resources. Its
new application in remote voice identification
systems, where a person is identified through a
telecommunications channel, seems promising.
For example, in mobile communications, voice
can be used to manage services, and the
introduction of voice identification helps protect
against fraud [
Voice identification is of particular
importance in the investigation of crimes, in
particular in the field of computer information,
and in the formation of the evidence base for such
an investigation. In these cases, it is often
necessary to identify an unknown voice
recording. Voice identification is an important
practical task when searching for a suspect based
on a voice recording in telecommunication
channels. Determining such characteristics of the
speaker’s voice as gender, age, nationality,
dialect, and emotional coloring of speech is also
important in the field of forensics and
antiterrorism. The identification results are
important in conducting phonoscopic
examinations, and in carrying out expert forensic
research based on the theory of forensic
identification [
Voice identification in real-world
environments faces the following serious
challenges. Firstly, such identification is subject
to all kinds of hardware distortions and noise
caused by the peculiarities of equipment and
devices for recording, processing, and storing
information. Secondly, external acoustic noise
inevitably superimposes the speech signal, which
can significantly distort individual informative
characteristics. Given this, identification systems
that have demonstrated fairly high efficiency in
laboratory conditions may show much lower
reliability when analyzing speech information
with external noise. Finally, in several tasks,
identification has to be performed in very difficult
conditions of overlapping voices of several
speakers, in particular, with similar acoustic
characteristics. It should be noted that there have
been virtually no studies of voice identification
capabilities for this most difficult case [
Voice identification involves a set of
technical, algorithmic, and mathematical
methods that cover all stages, from voice
recording to voice data classification. The
discussed difficulties and shortcomings lead to
the conclusion that further development of voice
identification systems requires the development
of new approaches aimed at processing large
arrays of experimental speech signals, their
effective analysis, and reliable classification. This
indicates the relevance of research on the
creation of new mathematical methods for
processing, analyzing, and classifying voice data
that would ensure the reliability and accuracy of
person identification [
Traditionally, the methods that provide the required level of classification reliability under given conditions are of practical interest for speech signal recognition. Until recently, the dominant approach to the construction of biometric voice identification devices was not to impose restrictions on the processing time of the speech signal, since the required recognition reliability was achieved by statistical processing of the results obtained, as well as by increasing the dimensionality of the recognition features, and as a rule, the process of extracting recognition features and their classification took place on the transmitting side of the communication channel.
However, in the case of remote voice
identification in modern mobile radio
communication systems, it is difficult to ensure
these conditions, since the identification of a
person is carried out on the receiving side, and
this, in turn, imposes certain factors of
computing power and the influence of noise in
the communication channel. An additional
requirement is often the need to make a
classification decision in a time-sensitive
environment [
In this case, it is necessary to move to other
methods that can provide the necessary
contrast of the speech signal in the formed
technologies in a remote
mode based on
modern mobile radio communication systems,
which will significantly expand the scope of this
type of technology. In this paper, we consider
the possibility of applying the theory of
timescale analysis to solve this problem [
In general, recognition is the process of assigning the object under study, in this case, a . . .. . .2 ],
is a -dimensional vector of observed values of features 1, 2, . . . , that reflect the most important properties of objects for recognition. The set of features , as a rule, is the same for all recognition classes 1, 2, . . . ,
Thus, we consider the task of recognizing the object under study belongs to one of a finite number of classes 1, 2, . . . , , which described by a set of features 1, 2, . . . , , which is the same for all classes. Differences are recognition features the specified conditions, namely, to ensure the quality of speech signal recognition features extraction influence noise
in by of under allow the the communication channel, which in turn use of voice identification
vector the will between classes will be manifested only in differences in the characteristics of features of different objects. Then, for any set of features 1, 2, . . . , , you can set rules according to which any two classes 1 and are assigned a = [ . . .
11 21 1 and 1 = . . [ 1 ] , which consists of parameters called interclass distances that express the degree of difference in the characteristics of recognition features [16].
An integral part of the speech signal recognition process is the definition of a set of features 1, 2, . . . , , i.e., the formation of recognition features in such a way as to ensure the required classification reliability with the minimum possible dimension .
By the considered approach to solving the problem of speech signal recognition, an important point is the choice of a method for forming recognition features. The use of approaches based on the traditional Fourier spectral-time analysis for this purpose is associated with certain difficulties. First, there are high requirements for the input speech signal stream in terms of signal-to-noise ratio. Secondly, the lack of classification reliability for multicomponent and low-stationary signals, such as speech signals, and thirdly, the need for a significant amount of implementations. The desire to overcome these limitations within the framework of traditional approaches of classical spectral signal processing leads to difficult-toimplement variants of speech signal recognition devices and solutions that are unacceptable for the conditions under consideration [17].
Thus, we formulate the research objective: to develop a method that allows the formation of contrasting recognition features for automatic remote identification of a person by voice under the conditions of restrictions on the duration of the processed realization at a signal-to-noise ratio of less than 20 dB under conditions of partial or complete a priori uncertainty about their structure.
Currently, analyzing methods
of decompose the input signal into a system of basis wavelets—functions, each of which is a shifted and scaled copy of the input (generating or mother wavelet). A characteristic property of wavelet functions (hereinafter referred to as wavelets) is the finite energy at their full localization in both frequency and time domains.
Thus, any sequence of discrete samples of the speech signal ( ) can be represented as an ordered set of coefficients of decomposition by a system of scaling functions and wavelet functions: ( ) = ∑ , , ( ) + ∑
∑ , , ( ), 2 − =1 =1 2 − =1 and
detailing coefficients [18]. defined analysis: where
is the number of decomposition levels, , and , are the approximating
Here, in (1) and (2) √2 is the normalizing factor, and = 0, ± 1, ± 2, . ..;
∈ .
In practice, to quickly calculate the values of wavelet coefficients , and ,
use a sequential separation scheme called the pyramid or Mallat algorithm, which is interpreted as a sequential two-band filtering of the input speech signal using cascaded low-pass (h) and high-pass (g) filter blocks (Fig. 1) [19]. In Fig. 1, for the wavelet coefficients , and , , the first index
corresponds to the number of the decomposition level, and the second index = 0,1, . . . , 2
− 1 corresponds to the ordinal value of the wavelet coefficient at the decomposition level . According to the theory of multiple-scale analysis, the values of , and , can be obtained based on the coefficients calculated at the previous stages of speech signal decomposition: , =
∑ −1, ℎ +2 , 1 √2 1 √2
, =
∑ −1, +2 , where ℎ and are sequences that define the characteristics of filters H and G at the level of wavelet decomposition [20].
The number of multiplication operations required to calculate all the coefficients of the discrete wavelet transform for the data set and the length of the vectors h and g equal to is 2
. The same number of operations is
required to recover or calculate all the spectral
components. So, to analyze a speech signal on a
wavelet basis, you need to perform 4
operations. The number of complex
multiplication operations for the fast Fourier
transform is 2 , which is comparable to
or even greater than in the case of the discrete
wavelet transform [
The interpretation of the coefficients of the discrete wavelet transform is somewhat more complicated than the Fourier coefficients. If the analyzed speech signal is sampled at a frequency of 8 kHz and consists of 256 samples, then the top frequency of the signal is 4 kHz.
Then the coefficients of the first level of decomposition (128) occupy the frequency band [2.0, 4.0] kHz. The second-level wavelet coefficients (64) are responsible for the [1.0, 2.0] kHz frequency band. They are displayed before the first level wavelet coefficients. The procedure is repeated until there is 1 wavelet coefficient and 1 scaling coefficient at level 9.
The total number of coefficients is (1+1+2+4+8+16+32+64+128) = 256. That is, the number of coefficients is equal to the number of samples in the input speech signal.
If the main energy of the signal was
concentrated near the frequency of 1.0 kHz,
then the second-level wavelet coefficients will
be more informative, and the first-level wavelet
coefficients can be neglected [
As a continuation of the development of the theory of multiple-scale analysis, it is proposed to improve the Mallat algorithm by additional processing of the high-frequency components of the pyramid of the analyzed speech signal.
Thus, in the improved algorithm, recursive filtering is applied to the coefficients , . This full decomposition algorithm is called wavelet packet decomposition. The decomposition scheme based on wavelet packets is shown in Fig. 2.
For the wavelet coefficients ( ) (Fig. 2), the index corresponds to the number of the decomposition level, the index corresponds to the number of the subband at the level , and = 0,1, . . . , 2 − 1 corresponds to the number of the wavelet coefficients at the level . In wavelet packages, several decomposition bases are used for complete decomposition, united by the image of nesting in each other, which gave the method its name. In general, each level of the hierarchy can use its specific basis. In contrast to the Mallat algorithm, the use of wavelet packets makes it possible to take into account the subtle structure of the analyzed speech signal process in a more comprehensive way.
Indeed, the absolute values of the
coefficients in the wavelet packet
decomposition are smaller than those of the
Mallat algorithm. Therefore, it can be argued
that the approximation with wavelet packets
has a much smaller error [
Since the wavelet basis is a complete decomposition basis, the wavelet coefficients contain individual characteristics of the input speech signal, determined by the properties of the basis functions to the same extent as the spectral components of the Fourier series.
Thus, any wavelet transform, including those based on the use of wavelet packets, allows you to uniquely represent a speech signal by an ordered set of its wavelet coefficients. It is possible to assume the possibility of using them as recognition features and thus put the calculation of coefficients based on wavelet packets based on the proposed method.
The method of forming speech signal
recognition features based on wavelet packets
is defined as follows. In the wavelet spectrum
formed based on wavelet packets, the power of
the calculated wavelet coefficients within each
subband of the decomposition is averaged. The
averaged coefficients are normalized and,
according to their place in the overall pyramid
of wavelet packets from left to right and from
top to bottom, converted into a vector of
recognition features. Thus, specific values of
the average power of the wavelet coefficients in
each subband of the decomposition will serve
as the primary features of speech signal
recognition. It should be noted that, in general,
the features obtained in this way will be
correlated, so it is advisable to apply an
additional decorrelation transformation to the
vector, which, by the way, will reduce the size
of the secondary recognition feature space
[
Consider the sequence of stages of the proposed method (Fig. 3). Z2,0 (i)
Z1,0 (i)
H G Z2,1 (i)
S (ti )
G
H Z2,2 (i)
Z1,1 (i)
G Z2,3 (i) H
G
H
G
H
G
H
G Zm,n (i) Zm,n (i) Zm,n (i) Zm,n (i) Zm,n (i) Zm,n (i) Zm,n (i) Zm,n (i)
Pm,n Pm,n Pm,n Pm,n Pm,n Pm,n Pm,n Pm,n Figure 3: Scheme of speech signal recognition features selection for biometric identification of a person Initially, the input sequence of discrete samples of the speech signal ( ) with length , a multiple of power 2, at = 0,1,2, . . . , (
− 1) is decomposed into ≤
2( )levels as a result of applying the wavelet packet algorithm. At the first level, the
input array ( ) is decomposed into two sets 1,0( ) and 1,1( ) by convolution ( ) with sequences {ℎ} and { }, which are determined by the characteristics of low H and high G frequency filters. At the 2nd level, the considered convolution procedures are repeated with each of the obtained subsets 1,0( ) and 1,1( ).
The process
of full
decomposition, called wavelet packetization,
involves steps similar to the first one [
The analytically considered procedures can be represented in general by the following expressions:
,2 ( ) = ∑ −1, ( )ℎ , ( ), ,2 +1( ) = ∑ −1, ( ) , ( ), where is 1 ≤
≤ , and 0 ≤ ≤ (2 −1 − 1).
At the first level of decomposition, the samples of the speech signal ( ) are used as 0,0( ) .
The values of the elements of the sequences {ℎ} and { } depend on the choice of the type of scaling function ( ) and wavelet function ( ) and, according to (1) and (2), are calculated as follows: ℎ , ( ) = 2− /2 (2− − ), , ( ) = 2− /2 (2− − ).
As a result of the
transformations performed during the
decomposition, the
sequence of samples of the speech signal ( )
is decomposed into = 2 ⋅ 2 − 1 sequences
(including the input one) of length /2 , each
of which represents one of the frequency
subbands of the input speech signal [
Different realizations of speech signals will have different energy distributions
over frequency subbands since their Fourier spectra will also be different. If you calculate the average power of the wavelet coefficients in each subband, the set of values obtained will reflect the wavelet content of the speech signal subbands, similar to the
frequency representation. Moreover, the transition to the average power will allow the use of relatively short input realizations for recognition, which is an important point in the operation of rapid analysis systems.
The bandwidth frequencies falling into each of the subbands will narrow with an increase in the number of the decomposition level, which follows from the wavelet
packet scheme (Fig. 2). The average powers of the wavelet coefficients in each subband, which are used as speech recognition features, are calculated according to the following expression: ̄ , = ∑ (( +1)⋅2 )−1( , ( ))2 = ⋅ /2 /2 .
(3)
power ̄0,0
realization ( ) [
To eliminate the sensitivity of the features to changes in the average power of the speech signal realization, the values of ̄ , obtained by (3) are normalized relative to the average calculated normalized values of ̄ , from left to right and from top to bottom. The number of the feature
expression = 2 is determined
according to the − 1 + and corresponds to the ordinal number of the component element of the vector
= { } .
An important point in implementing the method is the choice of the scaling function ( ) and the wavelet ( ). First, the size of the timefrequency window should be taken into account.
Second, the smoothness and symmetry of the underlying wavelet. Third, determine (set) the order of approximation. Correct selection of the wavelet basis for the speech signal significantly reduces the number of non-zero
wavelet
coefficients , ( ), which significantly reduces
the size of the recognition features and makes
them much more informative [
Practical experiments were conducted to investigate the contrast of the speech recognition feature vectors formed based on the proposed method. In particular, Figs. 4–5 show the feature vectors of the speech signal calculated in different wavelet decomposition bases.
Thus, in the first case (Fig. 4), a wavelet package based on the Haar basis was used to obtain wavelet coefficients, which provides a relatively coarse approximation of the speech signal, which accordingly affects the informativeness of the recognition features. In the second case (Fig. 5), the speech signal recognition features are calculated based on a smoother Meyer function, which makes the features more informative.
A comparative analysis of the results in Figs.
4–5 shows that when choosing a smoother basis
function, the number of yr values close to zero in
the feature vector = { } increases and the
informativeness of the decomposition increases,
unlike the Haar function, where we get less
informative recognition features. Thus, the use of
basic wavelet functions consistently in terms of
smoothness with the studied speech signal
allows us to reduce the size of recognition
features and increase their informativeness.
To confirm the hypothesis that it is expedient
to build speech signal recognition systems
based on wavelet packets using the values of
= { } obtained by expression (3) as
recognition features, we studied the developed
method of forming recognition features in
comparison with the approach proposed in
[15], which is based on the spectral
components of the classical harmonic Fourier
transform (Fig. 6).
The experiment used realizations of speech
signals with a duration of = 512 samples,
and the decomposition was performed at =
5 levels of decomposition. This approach
allowed us to obtain a feature vector = { }
of length = 32, where 16 wavelet coefficients
were averaged in each subband. As for the
recognition features based on the Fourier
transform, the spectrum was divided into 32
bands of 16 coefficients each [
RMSE is one of many metrics that are used to evaluate model performance. To calculate RMSE, square the number of detected errors and find the average value [30].
a b c signal recognition features using bases: a)
This experimental study is needed to compute an objective measure of the interclass distance RMSE of recognition features, i.e., the scatter of features when comparing different realizations of speech signals [31]. Comparative analysis of the existing and proposed methods
Harr, robustness of feature vectors formed based on wavelet packets for the Meyer, Haar basis and based on the Fourier energy spectrum, several experiments were conducted with the addition of white noise with a signal-to-noise ratio of 10 dB to the speech signal (noise power was measured in the analysis band) [32]. Fig. 8 shows all three feature vectors with the same noise power. c Figure 8: Thirty realizations of speech recognition features obtained at a signal-tonoise ratio of 10 dB based on bases: (a) Meyer, (b) Haar, and (c) Fourier Thus, it was possible to establish that at a signal-to-noise ratio of 10 dB, the features obtained based on the developed method have a very acceptable result, namely, a 1.6-2-fold increase in stability compared to the features obtained based on the traditional Fourier spectrum, where already at a signal-to-noise ratio of 20 dB the total deviation of the obtained features is unacceptable.
In this research, the task of extracting speech signal recognition features for voice identification of a person in a remote mode was solved, which imposes several restrictions, namely: (1) minimum processing time of the speech signal realization, since the required recognition reliability is achieved by statistical processing of the obtained results; (2) reduction of the dimensionality of recognition features, since the process of extracting recognition features and their classification occurs on the transmitting side of the communication channel, which in turn imposes certain factors of computing power and the influence of noise in the communication channel.
The studies have shown the ability of the developed method to form recognition features based on wavelet packets on the Meyer basis. The most important indicator of the effectiveness of the experiment is the increase in the contrast of recognition features, i.e., the increase in the interclass distance in the formed feature system for speech signals with a similar frequency-temporal structure. Even a visual analysis of the obtained values = { } (Figs. 7–8) reveals significant differences in the structure of the feature vectors formed by relatively short implementations, which proves the potential use of the presented method for speech signal recognition in rapid analysis systems. Since the recognition features are distributed according to normal law, the subsequent procedure for deciding whether speech signal realizations belong to a particular class is greatly simplified.
After analyzing the given conditions of the voice identification system, the question arose of developing a method for extracting speech signal recognition features that would provide more informative spectral characteristics of the speech signal, which would improve the efficiency of their further classification under the influence of noise.
This paper considers the possibility of applying the theory of scale-time analysis to solve this problem, namely, the development of a method for extracting recognition features based on the wavelet packet transform using the orthogonal basis wavelet Meyer function and subsequent averaging of wavelet coefficients that are in the frequency band of the corresponding wavelet packet. Experimental studies have shown the ability of the developed method to generate speech signal recognition features with a close frequency-temporal structure based on wavelet packets in the Meyer basis, namely, it was found that at a signal-to-noise ratio of 10 dB, the features obtained based on the developed method have a very acceptable result, namely, 1.6–2 times more robust to noise than the features obtained based on the traditional Fourier spectrum, where the total deviation of the root mean square error of the obtained features is unacceptable at a signalto-noise ratio of 20 dB.
Also, the analysis of the results shows that the contrast of the recognition features of test speech signals generated based on the developed method without the influence of noise is on average 3.8 times higher than that of the method using Fourier spectral coefficients.
The authors see the further direction of research in identifying the potential capabilities of the developed method of speech signal recognition in person identification under very difficult conditions of overlapping voices of several speakers, in particular with similar acoustic characteristics, as well as in selecting and justifying the criterion for implementing recognition procedures. It should be noted that there have been virtually no studies of voice identification capabilities for this most difficult case.