Speaker Recognition System Based on GMM Multivariate

Speaker Recognition System Based on GMM Multivariate
Piotr LENARCZYK, Zbigniew PIOTROWSKI
Military University of Technology, Faculty of Electronics
Speaker Recognition System Based on GMM Multivariate
Probability Distributions built-in a Digital Watermarking Token
Streszczenie. Przedstawiony poniżej artykuł opisuje system rozpoznawania mówcy na podstawie mowy ciągłej, wykorzystując wielowariancyjne
rozkłady prawdopodobieństwa GMM. Opisane zostały procesy ekstrakcji cech dystynktywnych głosu oraz tworzenia modeli statystycznych. Algorytm
został zaimplementowany w systemie Linux w celu poprawy funkcjonalności identyfikacji użytkownika Zaufanego Osobistego Terminalu PTT.
(System rozpoznawania mówcy na podstawie wielowariancyjnych rozkładów prawdopodobieństwa zaimplementowany w tokenie znaku
wodnego).
Abstract. The article describes a speaker recognition system based on continuous speech using GMM multivariate probability distributions. A
theoretical model of the system including the extraction of distinctive features and statistical modeling is described. The efficiency of the system
implemented in the Linux operating system was determined. The system is designed to support the functionality of the Personal Trusted Terminal
PTT in order to uniquely identify a subscriber using the device.
Słowa kluczowe: GMM, rozpoznawanie mówcy, PTT, biometria.
Keywords: GMM, speaker recognition, PTT, biometrics.
Introduction
Systems for identification of speaker biometric identity
allow for much more efficient information security
management since many services, such as authorization,
may be performed by means of individual and unique data
that is specific to a single person. A Trusted Personal
Terminal, PTT [1], [2] is a hardware digital watermarking
token developed for the purpose of identifying the
subscriber over open (unencrypted) communication links.
The digital watermark sent in speech signals transmitted
over the telecommunication link is inaudible, and represents
the subscriber's PIN. In order to enable the use of PTT only
to authenticated users a method of voice identification was
developed and described in this article. The subscriber
authentication model for open telephone links using the
PTT is a two-stage process. In the first stage a subscriber is
assigned to the PTT by voice authentication. In the second
stage a subscriber authenticated as above sends his or her
PIN using the PTT watermark token to the other side of the
link in order to identify their identity.
In general it can be said that the speech signal carries
two different types of information [3] - essentially it is
syntax, semantic and pragmatic in sense, but at the same
time, it is a rich source of information that clearly identifies
the speaker. A voice is a rich source of biometric data [4]
[5], due to the fact that it differs for each person in terms of
frequency, amplitude and phase in a sufficiently unique way
to distinguish various speakers. Both the vocal tract unique
in its construction and the signal stimulating this tract
generate the distinctive speech features assigned to a
particular person.
User recognition
Speaker recognition is the process of determining which
one of the registered users spoke a fragment of an
utterance. In these types of systems information about the
identity of the speaker is not given a 'priori and certain voice
characteristics are sought in the test utterance. Then they
are compared with a database of statistical parameters
stored for all users and the ones that best identify the
speaker are indicated. The implemented speaker
recognition system is of an open type, where the test
utterance belongs to one of the registered users. During
testing an adaptation coefficient is calculated for the set of
distinctive speech features extracted during the testing
process in relation to those calculated during learning
(training) and to the previously calculated UBM impostor
model. Calculated on the basis of statistical model
approximation of infinite number of users, it estimates a
model of a person from outside the set of authorized users.
The necessary condition is to calculate the UBM in
conditions of a set similar to the set of authorized users
(ratio of male to female, age, accent, etc.) where none of
the people used for estimating the UBM belongs to the set
of authorized users. Then the user with the highest
adaptation coefficient is recognized as the author of the test
utterance, or if all users have obtained similar results and
the testing of the impostor model on the basis of the test
utterance has achieved the highest score, the user will be
recognized as a new user (or impostor - for authentication
systems). A system of this type requires N+1 comparisons
for N users; well-designed speaker recognition algorithms
have a wide scope of application, thanks to low error rates
achieved [6], [7]. Speaker recognition algorithms can
operate on parameters calculated based on continuous
speech or isolated words (phrases). The described
algorithm uses MFCC continuous speech coefficients [8] for
testing the user model.
Theoretical model of the speaker recognition system
Two approaches to the problem of statistical modeling
are mainly used in the process of designing algorithms for
speaker recognition:
Hidden Markov Models (HMM) - this is the oldest approach
to this problem and it employs the modeling of each user as
a single Markov model.
Gaussian Mixture Models (GMM) - in this case the
dependence of modeling for each user is employed, in the
form of a sum of Gaussian probability distributions for
extracted distinctive speech feature vectors [7]
Fig. 1. Diagram of the training phase of the designed system
PRZEGLĄD ELEKTROTECHNICZNY, ISSN 0033-2097, R. 89 NR 2a/2013
59
Fig. 2. Diagram of the testing phase of the designed system
Speaker recognition system diagram
Figures 1 and 2 show diagrams of the two life cycle
phases of the system - training and testing. The training
phase is used to develop the model that statistically
estimates the extracted set of vectors identifying the
speaker identity, while the testing phase is used to indicate
the most probable attribution of the test utterance to one of
the available (authorized) user models, or to qualify it as an
impostor's utterance (or a new user).
In the training phase, following the extraction of vectors
of distinctive speech features,
VAD - Voice Activity
Detector is used. Its employment stems from the need to
reject such feature coefficients that contain unnecessary
information. In the real world, speech includes a lot of
pauses, respiratory breaks, environmental noise, etc. which
are undesirable in the process of GMM learning. For this
purpose, a two - distributive mixture model of multivariate
Gaussian probability function of the input multidimensional
set of extracted distinctive speech features is used. After
the two distributive classification process, there follows a
rejection of low-energy coefficients with corresponding
speech signal segments marked as unsuitable.
The next step consists in the Normalization of the set of
coefficients of distinctive speech features aimed at reducing
the environmental influence on the speech signal. In order
to remove the effect of using different microphones, A/D
and D/A converters, the telecommunication channel
characteristics, the coefficients are normalized in relation to
both the expectation and variance. A simple and effective
way is to calculate the average µ and expectation , taking
into account all the coefficients (assuming that for the
duration of the recording those are approximately constant).
The next step consists in the normalization of each
coefficient according to the formula:
(1)
Normalization relative to the mean is called Cepstral Mean
Substraction - CMS and historically it is the one used for the
longest period [9].
One of the basic ideas of speaker recognition systems is
the transformation of the vectors of the set of distinctive
speech feature coefficients into a more general model. This
process is called training or adaptation to training
coefficients. At least two models are developed in the
algorithm - that of a user and that of an impostor. They are
used in the testing process, where a statistic comparison is
used for the maximized probability of a statistical model of
the test utterance belonging to the model of particular users
or an impostor - on this basis a decision is made regarding
the classification of the user.
Figure 2 shows a simplified diagram of the testing stage
of the designed system. Initially distinctive features are
extracted from continuous speech of the test utterance and
then a statistical comparison of log likelihood coefficients is
performed, after calculation of testing speech samples
model adaptation to the user models and the impostor
60
model. The results obtained are normalized due to the fact
that, in spite of calculations performed earlier (sampling,
extraction and formation of a set of distinctive speech
feature vectors, GMM modeling, testing), which are
sufficient to develop a good system for user recognition,
studies show that the effectiveness of the algorithms can be
increased by normalizing the resulting coefficients. The last
- final stage of the testing phase is to decide on the basis of
the test utterance on the selection of the most similar user
model, or its rejection, if it will be most similar to impostor
model.
Extraction of distinctive speech features in speaker
recognition systems
An appropriate design of the extraction algorithm is one
of the most important elements of the speaker-recognition
system. The purpose of the extraction of distinctive speech
features is to achieve appropriate transformation of the
input speech signal into a sequence of speech features
vectors, where each of them represents information
contained in a frame. In other words, features extraction
transforms multielemental speech discrete signal into lowlevel features vector. The creation of the set of distinctive
speech feature vectors is performed based on MFCC
coefficients [8].
Fig. 3. Diagram of MFCC coefficient creation
Statistical modeling of users
In the case of multivariate Gaussian probability functions
the approach to speaker recognition is purely statistical. On
the basis of a set number of extracted distinctive speech
feature vectors (based on MFCC, calculated on the basis of
audio samples) a GMM model is created. This process
requires a set number of model parameters to be defined
(such as the amount of mixtures, the range of estimated
speech feature vectors, etc.).
The algorithm uses the fact that Gaussian distribution
can be extended to any dimensional variable in a domain
R , then it is called the multivariate Gaussian probability
function, for which it is possible to write:
(2)
lim
→
⋯
, ,
is the probability of encountering an infinitely small
quality coefficient around a multidimensional vector .
Twodimensional distribution is illustrated in figure 4 [10].
It is easy to notice that in the case of a single mixture
the set of distribution means is a vector, while the set of
variances is described as a n n dimensional covariance
matrix Σ . The use of the full covariance matrix makes it
possible to create a rotated distribution relative to the
expectations. Restricting
only to a diagonal matrix (this
type of matrix is called the variance matrix) prevents the
creation of rotated distributions, simplifying and shortening
the calculations performed. GMM models usually use only
diagonal variance matrices, as opposed to full covariance
PRZEGLĄD ELEKTROTECHNICZNY, ISSN 0033-2097, R. 89 NR 2a/2013
matrices, for describing multivariate Gaussian probability
density functions [6], [7], [11], [12], [13].
distinctive speech feature vectors). The Bayes Learning
Theory defines the relationship between Likelihood and A
Posteriori Probability:
Fig. 4. Two - dimensional GMM distribution
The mixture is a finite sum of N distributions:
(3)
∑
where p signifies single distributions and in the case of
GMM - Gaussian distributions. Weights w - determine the
impact of a particular single distribution on the whole
multivariate distribution. In accordance with probability
theory for 0 p , p
1 the sum of the weights must be 1
(otherwise p x it is not a probability distribution). The use
of GMM in speaker recognition algorithms requires at least
two algorithms.
First, it is necessary to define the way of obtaining each
values of the set of distributions first moments, variances
and weights to statistically model multi-dimensional value
distributions of distinctive speech feature vectors. They are
obtained through the EM algorithm (Expectation
Maximization - described later in this article) which
calculates 3N parameters for the mixture of
distribution,
where each distribution is described by three parameters:
expectation μ, variance
and weight .
Secondly, it is necessary to define the method for
evaluating distinctive speech feature vectors for particular
audio segments in order to identify a particular speaker.
Due to the fact that the output result of GMM is a probability
value, it is necessary to determine the decision-making
method. The simplest and most widely used method is to
employ thresholding which for applications in speaker
recognition can be described as a process of assigning any
value to one of two classes.
Thresholding is a single estimated value independent of
the number of users. Usually, [7], [11], [13] a constant
threshold is used, user independent, due to the fact that in
this case there is no need to compare the statistical
assessment of the distinctive speech feature vectors for
each user (from a set of authorized users) with all other
estimated models for other users.
Likelihood and A Posteriori Probability
During the GMM training and testing process two kinds
of probabilities are employed, which represent two different
data properties. The first is Likelihood , |
- the
probability that the data was generated under the
conditions of the Hypothesis
. This hypothesis states that
the tested user is an authorized user and assumes in the
query that the user is authorized through an assigned
authorized user model
from among the available
authorized user models. In speaker recognition algorithms
this probability is described through an inverse relationship
|
- the resulting A Posteriori Probability checking
whether the tested user's identity is the true identity (from
set of authorized users and impostor model), taking into
account the tested segment of audio data (extracted
|
|
(4)
Probability p H
is A Priori Probability which
determines the probability of the speaker's tested identity
being the true identity. This means that the system is to set
higher probability values for "typical" users and lower for
users with more sparse values distribution of distinctive
speech feature vectors. Due to the fact that speaker
recognition systems are designed for a large number of
users, it is assumed that this probability is constant for all
users, and as it is easy to notice, with the multiplicity of the
set of authorized users going towards infinity the A Priori
Probability goes towards zero. Therefore probability
is ignored - taking it into account for a finite set of users is
employed merely for scaling. Likelihood p x is the
probability of finding features in a set of all features. That
is why typical values of distinctive speech features have an
insignificant influence over GMM, contrasted with the rare
ones. The most commonly used Neyman - Pearson Lemm's
[14] probability
is described later in this article.
EM algorithm
The most often used method for training GMM models is
the EM algorithm [15]. The issue of training consists in
using a finite number of samples in order to determine
optimal parameters for modeling. In the case of speaker
recognition the samples are usually segmented into parts in
which the user utters certain word sequences. The training
of the model is done through iteration within a segment and
finding gradually better values of the parameters (sets:
weights, averages and variances). Due to the fact that the
EM algorithm is not able to achieve final result, a
Maximization step must be performed. In speaker
recognition applications this is done through Likelihood, or
A Posteriori Probability. Therefore, the EM algorithm is one
way to find Maximum Likelihood – ML or Maximum A
Posteriori Probability - MAP.
Let us assume that
describes the parameters of the
model, ∗ describes the optimal parameters, ′ describes
the new value
after a single optimizing iteration.
Therefore, the EM algorithm iteratively finds ′ which
maximizes the probability of finding the analyzed variables
values, taking into account the current values of model
parameters, so:
, |
(5)
| ,
where
is the available data, Y is a random variable
describing unknown (hidden) data. The expectation is
usually described as Q θ and calculating its value is the
first step in each iteration.
EM algorithm used in multivariate Gaussian probability
functions.
Calculating the optimal GMM model parameters through
maximization ML [16], as well as MAP [7] is complicated,
but can be summarized as the most significant formulas
described below.
Calculating the optimal GMM model parameters with
Probability Maximization ML is limited to three basic
formulas for updating parameters of the mixtures model
[16]:
(6)
(7)
PRZEGLĄD ELEKTROTECHNICZNY, ISSN 0033-2097, R. 89 NR 2a/2013
,….,
;
,…,
∑
;
,…,
| ,
61
∑
(8)
|
∑
∑
′
(9)
A Posteriori Probability is described by the formula:
,
,
|
|
∑
(20)
,
,
|
where w is the weight for the i- distribution and p i|x , θ is
the probability that x was generated from this i-distribution.
Therefore, the weighted average is:
|
,
(10)
∑
|,
Calculating the optimal GMM model parameters with
Maximum A Posteriori Probability - MAP is done using the
formulas [5]:
(11)
1
(12)
(13)
1
(14)
∑
′
,
∑
(15)
(16)
| ,
∑
| ,
where N - is the number of distributions, r and α are
constants, described later in this article. The algorithm
consists in a simple repetition of three formulas, taking into
account the condition of convergence, or a defined number
of repetitions. The algorithm of Maximization of A Posteriori
Probability is especially sensitive to model overtraining,
where during statistical modeling, when multivariate
Gaussian probability functions are used, overly detailed
data will be taken into account and the model will not fulfill
its role properly.
|
|
where it was mentioned that
in this form is rarely used
due to the fact that it requires significantly more information
than needed (information about all possible human voices).
However, based on the fact that the appropriate number of
users (non-authorized) is used as part of training the
Impostor Model it is possible to use a simple probability
instead of log p H |x, θ .
calculation logp H |x, θ
Formalizing: p H
1 and the final form of the probability:
(21)
|
|
On this basis the decision process is performed which
consists in making a decision regarding the choice of the
most similar user model, or its rejection, if it will be most
similar to the impostor model.
Description of a practical implementation of the
speaker recognition system
The theoretical considerations described earlier were
implemented in practice in the form of a C++ application, in
the QT environment (version 4.7.4), on the Linux operating
system (Mandriva, kernel version 2.6.39.4-5.1) with the
purpose of porting the algorithm to the PTT hardware
platform. A simplified diagram of the implemented speaker
recognition system is illustrated in figure 5.
Testing the speaker's set of distinctive speech feature
vectors
Once GMM modeling is finished for both the impostor
and authorized user, the next step is the testing stage
described in figure 2. In this stage the system must be able
to recognize the speaker identity and reach a decision at
the beginning. In the case of testing the simplest answer
could be a single test coefficient for a single voice and
model segment. A more sophisticated response is a vector
containing test coefficients responsible for different voice
characteristics. The most often used are the Likelihood
Ratio - LR and the Log Likelihood Ratio - LLR as defined in
[17]. This is the optimum test coefficient minimizing the
number of errors of the FN type - False Negatives, that is,
taking wrong decision on a negative decision given for the
tested utterance. It is described by the formula:
| ,
(17)
(18)
| ,
log
log
| ,
log
| ,
where
| ,
is the probability that H is not the valid
hypothesis for x (M is not an authorized user for his tested
utterance). A useful approximation for the LLR testing
coefficient: p H |x is the most widely used, due to the fact
that it is possible to skip all users except the tested
in
calculations for the GMM model, which in practice means
skipping consideration for all possible human voices. In this
case the fact that H is not part of
is used. Thanks to
this approximation calculations are significantly simplified:
(19)
62
log
| ,
| ,
Fig. 5. Simplified diagram of the implemented speaker recognition
system
On the basis of the practically implemented speaker
recognition system, the effectiveness of the algorithm was
tested (using FAR - False Accept Rate and FRR - False
Rejection Rate coefficients in order to calculate the EER Equal Error Rate coefficient of the implemented system),
under the following conditions: 50 users (40 men and 10
women) who trained the UBM model, 40 users tested (30
men and 10 women), number of mixtures of GMM UBM
mixtures 512, number of mixtures of GMM user models
512, total duration of utterances of users training the UBM
PRZEGLĄD ELEKTROTECHNICZNY, ISSN 0033-2097, R. 89 NR 2a/2013
model 2.5 hours, duration of training utterances of a single
user ~ 180s (numbers repeated in sequence), testing
utterance time ~5s (random words found in a dictionary),
number of testing utterances 1480, recording sample rate
44100Hz, bits per sample 16, number of channels 2 ,
number of microphones 7, size of the training and testing
dictionary 10 words (digits), resulting efficiency of the
algorithm is: EER = 2.13%.
The high accuracy of the speaker recognition results
stems mainly from a small number of people tested and
high quality recordings and from the fact that the set of
words in dictionary was low (10 elements).
Summary
The article presents an algorithm for biometric speaker
identity recognition based on continuous speech using
multivariate probability distributions. The system is
designed to support the functionality of the Personal
Trusted Terminal PTT in order to uniquely identify a
subscriber using the device. The theoretical model of the
statistical modeling algorithm has been accurately
described, its practical implementation and test results were
shown.
Acknowledgements
This paper has been financed from science funds
granted within the years 2010-2012 as a research project of
the Polish National Centre for Research and Development
of the Republic of Poland No. 0181/R/T00/2010/12.
REFERENCES
[1] Piotrowski Z., Zagoździński L., Gajewski P., Nowosielski
L.,Handset with hidden authorization function, European DSP
Education & Research Symposium EDERS (2008),
Proceedings, 201-205, Texas Instruments
[2] Piotrowski Z., The NNC System and its components in the age
of Information Warfare, Safety and Security Engineering III,
SAFE III, WIT Press, Southampton, Boston,(2009), 301-309
[3] Davis, S., Merlmestein, P., Comparison of Parametric
Representations for Monosyllabic Word Recognition in
Continuously Spoken Sentences, IEEE Trans. on ASSP (1980),
357-366
[4] Neiberg D., Text Independent Speaker Verication Using
Adapted Gaussian Mixture Models CTT (2001)
[5] Joseph P., Campbell, Jr., Speaker recognition, PII: S 00189219(97)06947-8
[6] Reynolds D.A., Rose R.C., Robust text-independent speaker
identification using Gaussian mixture speaker models, IEEE
Trans. Speech Audio Process, 3 (1995), 72–83
[7] Reynolds D.A., Speaker identification and verification using
Gaussian mixture speaker models, Speech Commun. 17
(1995), 91–108
[8] Rabiner L., Juang B.H., Fundamentals of Speech Recognition,
Prentice-Hall (1993)
[9] Markel J., Oshika B., Gray Jr. A., Long-term feature averaging
for speaker reognition. ZEEE Transactions on Acoustics,
Speech, and Signal Processing, August (1977),54-61
[10] McLachlan G., Peel D., Finite Mixture Models. Hoboken, NJ:
John Wiley & Sons, Inc., (2000)
[11] Reynolds D.A., Automatic speaker recognition using Gaussian
mixture speaker models, Lincoln Lab. J, 8 (1996), 173–192
[12] Reynolds D.A., Comparison of Background Normalization
Methods
for
Text-Independent
Speaker
Verification,
Proceedings of Eurospeech, (1997), 963–966
[13] Reynolds D.A., Quatieri T.F., Dunn R.B, Speaker Verification
Using Adapted Gaussian Mixture Models, Digital Signal
Processing, (2000)
[14] Jiang H., Confidence Measures For Speech Recognition,
Survey A., Speech Communication, Volume 45 No. 4 (2005),
455–470
[15] Dempster A.P., Laird N.M., Rubin D.B., Maximum-Likelihood
From Incomplete Data via the EM Algorithm, Journal of the
Royal Statistical Society, Ser. B., 39 (1977)
[16] Bilmes, J., Gentle A., Tutorial of the EM Algorithm and its
Application to Parameter Estimation for Gaussian Mixture and
Hidden Markov Models, International Computer Science
Institute, (1998)
[17] Jiang, H., Confidence Measures For Speech Recognition,
Survey A., Speech Communication, Volume 45 No. 4(2005),
455–470
Authors: M.Sc. Piotr Lenarczyk, Military University of Technology,
Faculty of Electronics, Warsaw, Poland, Tel. +48512127726, Fax.
+4822-683-90-3, E-mail [email protected]; PhD. Zbigniew
Piotrowski, Military University of Technology Faculty of Electronics,
Warsaw, Poland, Tel. +4822-683-97-99, Fax. +4822-683-90-3, Email [email protected]
PRZEGLĄD ELEKTROTECHNICZNY, ISSN 0033-2097, R. 89 NR 2a/2013
63
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement