Chapter 2

Gaussian Mixture Models

Abstract In this chapter we first introduce the basic concepts of random variables
and the associated distributions. These concepts are then applied to Gaussian random
variables and mixture-of-Gaussian random variables. Both scalar and vector-valued
cases are discussed and the probability density functions for these random variables
are given with their parameters specified. This introduction leads to the Gaussian
mixture model (GMM) when the distribution of mixture-of-Gaussian random vari-
ables is used to fit the real-world data such as speech features. The GMM as a
statistical model for Fourier-spectrum-based speech features plays an important role
in acoustic modeling of conventional speech recognition systems. We discuss some
key advantages of GMMs in acoustic modeling, among which is the easy way of
using them to fit the data of a wide range of speech features using the EM algorithm.
We describe the principle of maximum likelihood and the related EM algorithm for
parameter estimation of the GMM in some detail as it is still a widely used method
in speech recognition. We finally discuss a serious weakness of using GMMs in
acoustic modeling for speech recognition, motivating new models and methods that
form the bulk part of this book.

2.1 Random Variables

The most basic concept in probability theory and in statistics is the random variable.
A scalar random variable is a real-valued function or variable, which takes its value
based on the outcome of a random experiment. A vector-valued random variable is
a set of scalar random variables, which may either be related to or be independent
of each other. Since the experiment is random, the value assumed by the random
variable is random as well. A random variable can be understood as a mapping from
a random experiment to a variable. Depending on the nature of the experiment and
of the design of the mapping, a random variable can take either discrete values,
continuous values, or a mix of discrete and continuous values. We, hence, see the
names of discrete random variable, continuous random variable, or hybrid random
variable. All possible values, which may be assumed by a random variable, are

© Springer-Verlag London 2015 13

D. Yu and L. Deng, Automatic Speech Recognition,
Signals and Communication Technology, DOI 10.1007/978-1-4471-5779-3_2
14 2 Gaussian Mixture Models

sometimes called its domain. In this as well as a few other later chapters, we use
the same notations to describe random variables and other concepts as those adopted
in [16].
The fundamental characterization of a continuous-valued random variable, x, is
its distribution or the probability density function (PDF), denoted generally by p(x).
The PDF for a continuous random variable at x = a is defined by

. P(a − Δa < x ≤ a)
p(a) = lim ≥0 (2.1)
Δa→0 Δa

where P(·) denotes the probability of the event.

The cumulative distribution function of a continuous random variable x evaluated
at x = a is defined by

a
.
P(a) = P(x ≤ a) = p(x)d x. (2.2)
−∞

A PDF has to satisfy the property of normalization:

∞
P(x ≤ ∞) = p(x)d x = 1. (2.3)
−∞

If the normalization property is not held, we sometimes call the PDF an improper
density or unnormalized distribution.
For a continuous random vector x = (x1 , x2 , . . . , x D )T ∈ R D, we can similarly
define their joint PDF of p(x1 , x2 , . . . , x D ). Further, a marginal PDF for each of the
random variable xi in the random vector x is defined by
  
.
p(xi ) = ... p(x1 , . . . , x D ) d x1 . . . d xi−1 d xi+1 . . . d x D . (2.4)
all x j : x j =xi

It has the same properties as the PDF for a scalar random variable.

2.2 Gaussian and Gaussian-Mixture Random Variables

A scalar continuous random variable x is normally or Gaussian distributed if its

PDF is
2.2 Gaussian and Gaussian-Mixture Random Variables 15
   
1 1 x −μ 2 .
p(x) = exp − = N (x; μ, σ 2 ),
(2π )1/2 σ 2 σ
(−∞ < x < ∞; σ > 0) (2.5)

An equivalent notation for the above is

x ∼ N (μ, σ 2 ),

denoting that random variable x obeys a normal distribution with mean μ and variance
σ 2 . With the use of the precision parameter r, a Gaussian PDF can also be written as
  r
r
p(x) = exp − (x − μ)2 . (2.6)
2π 2

It is a simple exercise to show that for a Gaussian random variable x, E(x) =

μ, var (x) = σ 2 = r −1 .
The normal random vector x = (x1 , x2 , . . . , x D )T , also called multivariate or
vector-valued Gaussian random variable, is defined by the following joint PDF:


1 1 T −1 .
p(x) = exp − (x − μ) Σ (x − μ) = N (x; μ, Σ) (2.7)
(2π ) D/2 |Σ|1/2 2

An equivalent notation is x ∼ N (μ ∈ R D , Σ ∈ R D×D ). It is also straight

forward to show that for a multivariate Gaussian random variable, the expectation
and covariance matrix are given by E(x) = μ; E[(x − x)(x − x)T ] = Σ.
The Gaussian distribution is commonly used in many engineering and science
disciplines including speech recognition. The popularity arises not only from its
highly desirable computational properties, but also from its ability to approximate
many naturally occurring real-world data, thanks to the law of large numbers.
Let us now move to discuss the Gaussian-mixture random variable with the dis-
tribution called mixture of Gaussians. A scalar continuous random variable x has a
Gaussian-mixture distribution if its PDF is specified by
   

M
cm 1 x − μm 2
p(x) = exp − (2.8)
(2π )1/2 σm 2 σm
m=1

M
= cm N (x; μm , σm2 ) (−∞ < x < ∞; σm > 0; cm > 0)
m=1

M
where the positive mixture weights sum to unity: m=1 cm = 1.
The most obvious property of Gaussian mixture distribution is its multimodal
one (M > 1 in Eq. 2.8), in contrast to the unimodal property of the Gaussian
distribution where M = 1. This makes it possible for a mixture Gaussian distribution
16 2 Gaussian Mixture Models

to adequately describe many types of physical data (including speech data) exhibiting
multimodality poorly suited for a single Gaussian distribution. The multimodality
in data may come from multiple underlying causes each being responsible for one
particular mixture component in the distribution. If such causes are identified, then
the mixture distribution can be decomposed into a set of cause-dependent or context-
dependent component distributions.
It is easy to show that the expectation
M of a random variable x with the mixture
Gaussian PDF of Eq. 2.8 is E(x) = m=1 cm μm . But unlike a (uni-modal) Gaussian
distribution, this simple summary statistic is not very informative unless all the
component means, μm , m = 1, . . . , M, in the Gaussian-mixture distribution are
close to each other.
The multivariate generalization of the mixture Gaussian distribution has the joint
PDF of


M

cm 1 T −1
p(x) = exp − (x − μm ) Σ m (x − μm )
(2π ) D/2 |Σ m |1/2 2
m=1

M
= cm N (x; μm , Σ m ), (cm > 0). (2.9)
m=1

The use of this multivariate mixture Gaussian distribution has been one key factor
contributing to improved performance of many speech recognition systems (prior to
the rise of deep learning); e.g., [14, 23, 24, 27]. In most applications, the number of
mixture components, M, is chosen a priori according to the nature of the problem,
although attempts have been made to sidestep such an often difficult problem of
finding the “right” number; e.g., [31].
In using the multivariate mixture Gaussian distribution of Eq. 2.8, if the variable
x’s dimensionality, D, is large (say, 40, for speech recognition problems), then the
use of full (nondiagonal) covariance matrices (Σ m ) would involve a large number
of parameters (on the order of M × D 2 ). To reduce the number of parameters, one
can opt to use diagonal covariance matrices for Σ m . Alternatively, when M is large,
one can also constrain all covariance matrices to be the same; i.e., “tying” Σ m for
all mixture components, m. An additional advantage of using diagonal covariance
matrices is significant simplification of computations needed for the applications of
the Gaussian-mixture distributions. Reducing full covariance matrices to diagonal
ones may have seemed to impose uncorrelatedness among data vector components.
This has been misleading, however, since a mixture of Gaussians each with a diagonal
covariance matrix can at least effectively describe the correlations modeled by one
Gaussian with a full covariance matrix.
2.3 Parameter Estimation 17

2.3 Parameter Estimation

The Gaussian-mixture distributions we just discussed contain a set ofparameters. In

the multivariate case of Eq. 2.8, the parameter set consists of Θ = cm, μm , Σ m .
The parameter estimation problem, also called learning, is to determine the values of
these parameters from a set of data typically assumed to be drawn from the Gaussian-
mixture distribution.
It is common to think of Gaussian mixture modeling and the related parameter
estimation as a missing data problem. To understand this, let us assume that the data
points under consideration have “membership,” or the component of the mixture, in
one of the individual Gaussian distributions we are using to model the data. At the
start, this membership is unknown, or missing. The task of parameter estimation is
to learn appropriate parameters for the distribution, with the connection to the data
points being represented as their membership in the individual Gaussian distributions.
Here, we focus on maximum likelihood methods for parameter estimation of the
Gaussian-mixture distribution, and the expectation maximization (EM) algorithm
in particular. The EM algorithm is the most popular technique used to estimate the
parameters of a mixture given a fixed number of mixture components, and it can
be used to compute the parameters of any parametric mixture distribution. It is an
iterative algorithm with two steps: an expectation or E-step and a maximization or
M-step. We will cover the general statistical formulation of the EM algorithm, based
on [5], in more detail in Chap. 3 on hidden Markov models, and here we only discuss
its practical use for the parameter estimation problem related to the Gaussian mixture
distribution.
The EM algorithm is of particular appeal for the Gaussian mixture distribution
as the main topic of this chapter, where closed-form expressions in the M-step are
available as expressed in the following iterative fashion1 :

1  ( j)
N
( j+1)
cm = h m (t), (2.10)
N
t=1
N ( j) (t)
( j+1) t=1 h m (t)x
μm = N ( j)
, (2.11)
t=1 h m (t)
N ( j) ( j) ( j)
( j+1) h m (t)[x(t) − μm ][x (t) − μm ]T
Σm = t=1 N ( j) , (2.12)
t=1 h m (t)

where the posterior probabilities (also called the membership responsibilities) com-
puted from the E-step are given by

1 Detailed derivation of these formulae can be found in [1], which we omit here. Related derivations
for similar but more general models can be found in [2, 3, 6, 15, 18].
18 2 Gaussian Mixture Models

( j) ( j) ( j)
( j) cm N (x(t) ; μm , Σ m )
h m (t) = n ( j) ( j) ( j)
(2.13)
(t)
i=1 ci N (x ; μi , Σ i ).

That is, on the basis of the current (denoted by superscript j above) estimate for the
parameters, the conditional probability for a given observation x(t) being generated
from mixture component m is determined for each data sample point at t = 1, . . . , N ,
where N is the sample size. The parameters are then updated such that the new compo-
nent weights correspond to the average conditional probability and each component
mean and covariance is the component specific weighted average of the mean and
covariance of the entire sample set.
It has been well established that each successive EM iteration will not decrease
the likelihood, a property not shared by most other gradient based maximization
techniques. Further, the EM algorithm naturally embeds within it constraints on the
probability vector, and for sufficiently large sample sizes positive definiteness of the
covariance iterates. This is a key advantage since explicitly constrained methods incur
extra computational costs to check and maintain appropriate values. Theoretically,
the EM algorithm is a first-order one and as such converges slowly to a fixed-point
solution. However, convergence in likelihood is rapid even if convergence in the
parameter values themselves is not. Another disadvantage of the EM algorithm is its
propensity to spuriously identify local maxima and its sensitivity to initial values.
These problems can be addressed by evaluating EM at several initial points in the
parameter space although this may become computationally costly. Another popular
approach to address these issues is to start with one Gaussian component and split
the Gaussian components after each epoch.
In addition to the EM algorithm discussed above for parameter estimation that
is rested on maximum likelihood or data fitting, other types of estimation aimed
to perform discriminative estimation or learning have been developed for Gaussian
or Gaussian mixtures, as special cases of the related but more general statistical
models such as the Gaussian HMM and its Gaussian-mixture counterpart; e.g.,
[22, 25, 26, 33].

2.4 Mixture of Gaussians as a Model for the Distribution of

Speech Features

When speech waveforms are processed into compressed (e.g., by taking logarithm
of) short-time Fourier transform magnitudes or related cepstra, the Gaussian-mixture
distribution discussed above is shown to be quite appropriate to fit such speech
features when the information about the temporal order is discarded. That is, one can
use the Gaussian-mixture distribution as a model to represent frame-based speech
features. We use the Gaussian mixture model (GMM) to refer to the use of the
Gaussian-mixture distribution for representing the data distribution. In this case and
in the remainder of this book, we generally use model or computational model
2.4 Mixture of Gaussians as a Model for the Distribution of Speech Features 19

to refer to a form of mathematical abstraction of aspects of some realistic physi-

cal process (such as the human speech process), following the guiding principles
detailed in [9]. Such models are established often with necessary simplification and
approximation aimed at mathematical or computational tractability. The tractability
is crucial in making the mathematical abstraction amenable to computer or algorith-
mic implementation for practical engineering applications (such as speech analysis
and recognition).
Both inside and outside the speech recognition domain, the GMM is commonly
used for modeling the data and for statistical classification. GMMs are well known
for their ability to represent arbitrarily complex distributions with multiple modes.
GMM-based classifiers are highly effective with widespread use in speech research,
primarily for speaker recognition, denoising speech features, and speech recognition.
For speaker recognition, the GMM is directly used as a universal background model
(UBM) for the speech feature distribution pooled from all speakers [4, 28, 32, 34]. In
speech feature denoising or noise tracking applications, the GMM is used in a similar
way and as a prior distribution [10–13, 19, 21]. In speech recognition applications,
the GMM is integrated into the doubly stochastic model of HMM as its output
distribution conditioned on a state, a topic which will be discussed in a great detail
in Chap. 3.
When speech sequence information is taken into account, the GMM is no longer
a good model as it contains no sequence information. A class of more general mod-
els, called the hidden Markov models (HMM) to be discussed in Chap. 3, captures
the sequence information. Given a fixed state of the HMM, the GMM remains a
reasonably good model for the PDF of speech feature vectors allocated to the state.
GMMs have several distinct advantages that make them suitable for modeling
the PDFs over speech feature vectors associated with each state of an HMM. With
enough components, they can model PDFs to any required level of accuracy, and
they are easy to fit to data using the EM algorithm described in Sect. 2.3. A huge
amount of research has gone into finding ways of constraining GMMs to increase
their evaluation speed and to optimize the tradeoff between their flexibility and the
amount of training data required to avoid overfitting. This includes the development
of parameter-tied or semi-tied GMMs and subspace GMMs.
Beyond the use of the EM algorithm for parameter estimation of the GMM
parameters, the speech recognition accuracy obtained by a GMM-based system
(which is interfaced with the HMM) has been drastically improved if the GMM
parameters are discriminatively learned after they have been generatively trained by
EM to maximize its probability of generating the observed speech features in the
training data. This is especially true if the discriminative objective function used
for training is closely related to the error rate on phones, words, or sentences. The
accuracy can also be improved by augmenting (or concatenating) the input speech
features with tandem or bottleneck features generated using neural networks, which
we will discuss in a later chapter. GMMs had been very successful in modeling
speech features and in acoustic modeling for speech recognition for many years
(until around year 2010–2011 when deep neural networks were shown to outper-
form the GMMs).
20 2 Gaussian Mixture Models

Despite all their advantages, GMMs have a serious shortcoming. That is, GMMs
are statistically inefficient for modeling data that lie on or near a nonlinear mani-
fold in the data space. For example, modeling the set of points that lie very close to
the surface of a sphere only requires a few parameters using an appropriate model
class, but it requires a very large number of diagonal Gaussians or a fairly large
number of full-covariance Gaussians. It is well-known that speech is produced
by modulating a relatively small number of parameters of a dynamical system
[7, 8, 17, 20, 29, 30]. This suggests that the true underlying structure of speech
is of a much lower dimension than is immediately apparent in a window that
contains hundreds of coefficients. Therefore, other types of model, which can
capture better properties of speech features, are expected to work better than GMMs
for acoustic modeling of speech. In particular, the new models should more effectively
exploit information embedded in a large window of frames of speech features than
GMMs. We will return to this important problem of characterizing speech features
after discussing a model, the HMM, for characterizing temporal properties of speech
in the next chapter.

References

1. Bilmes, J.: A gentle tutorial of the EM algorithm and its application to parameter estimation
for Gaussian mixture and hidden Markov models. Technical Report, TR-97-021, ICSI (1997)
2. Bilmes, J.: What HMMs can do. IEICE Trans. Inf. Syst. E89-D(3), 869–891 (2006)
3. Bishop, C.: Pattern Recognition and Machine Learning. Springer, Heidelberg (2006)
4. Dehak, N., Kenny, P., Dehak, R., Dumouchel, P., Ouellet, P.: Front-end factor analysis for
speaker verification. IEEE Trans. Audio, Speech Lang. Process. 19(4), 788–798 (2011)
5. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum-likelihood from incomplete data via the
EM algorithm. J. Royal Stat. Soc. Ser. B. 39, 1–38 (1977)
6. Deng, L.: A generalized hidden markov model with state-conditioned trend functions of time
for the speech signal. Signal Process. 27(1), 65–78 (1992)
7. Deng, L.: Computational models for speech production. In: Computational Models of Speech
Pattern Processing, pp. 199–213. Springer, New York (1999)
8. Deng, L.: Switching dynamic system models for speech articulation and acoustics. In: Math-
ematical Foundations of Speech and Language Processing, pp. 115–134. Springer, New York
(2003)
9. Deng, L.: Dynamic Speech Models—Theory, Algorithm, and Applications. Morgan and Clay-
pool, New York (2006)
10. Deng, L., Acero, A., Plumpe, M., Huang, X.: Large vocabulary speech recognition under ad-
verse acoustic environment. In: Proceedings of International Conference on Spoken Language
Processing (ICSLP), pp. 806–809 (2000)
11. Deng, L., Droppo, J.: A. Acero: recursive estimation of nonstationary noise using iterative
stochastic approximation for robust speech recognition. IEEE Trans. Speech Audio Process.
11, 568–580 (2003)
12. Deng, L., Droppo, J., Acero, A.: A Bayesian approach to speech feature enhancement using
the dynamic cepstral prior. In: Proceedings of International Conference on Acoustics, Speech
and Signal Processing (ICASSP), vol. 1, pp. I-829–I-832 (2002)
13. Deng, L., Droppo, J., Acero, A.: Enhancement of log mel power spectra of speech using a
phase-sensitive model of the acoustic environment and sequential estimation of the corrupting
noise. IEEE Trans. Speech Audio Process. 12(2), 133–143 (2004)
References 21

14. Deng, L., Kenny, P., Lennig, M., Gupta, V., Seitz, F., Mermelsten, P.: Phonemic hidden markov
models with continuous mixture output densities for large vocabulary word recognition. IEEE
Trans. Acoust, Speech Signal Process. 39(7), 1677–1681 (1991)
15. Deng, L., Mark, J.: Parameter estimation for markov modulated poisson processes via the em
algorithm with time discretization. In: Telecommunication Systems (1993)
16. Deng, L., O’Shaughnessy, D.: Speech Processing—A Dynamic and Optimization-Oriented
Approach. Marcel Dekker Inc, New York (2003)
17. Deng, L., Ramsay, G., Sun, D.: Production models as a structural basis for automatic speech
recognition. Speech Commun. 33(2–3), 93–111 (1997)
18. Deng, L., Rathinavelu, C.: A Markov model containing state-conditioned second-order non-
stationarity: application to speech recognition. Comput. Speech Lang. 9(1), 63–86 (1995)
19. Deng, L., Wang, K., Acero, A., Hon, H., Droppo, J., Boulis, C., Wang, Y., Jacoby, D., Mahajan,
M., Chelba, C., Huang, X.: Distributed speech processing in mipad’s multimodal user interface.
IEEE Trans. Audio Speech Lang. Process. 20(9), 2409–2419 (2012)
20. Divenyi, P., Greenberg, S., Meyer, G.: Dynamics of Speech Production and Perception. IOS
Press, Washington (2006)
21. Frey, B., Deng, L., Acero, A., Kristjansson, T.: Algonquin: iterating laplaces method to remove
multiple types of acoustic distortion for robust speech recognition. In: Proceedings of European
Conference on Speech Communication and Technology (EUROSPEECH) (2000)
22. He, X., Deng, L.: Discriminative Learning for Speech Recognition: Theory and Practice. Mor-
gan and Claypool, New York (2008)
23. Huang, X., Acero, A., Hon, H.W., et al.: Spoken Language Processing. Prentice Hall, Engle-
wood Cliffs (2001)
24. Huang, X., Deng, L.: An overview of modern speech recognition. In: Indurkhya, N.,
Damerau, F.J. (eds.) Handbook of Natural Language Processing, 2nd edn. CRC Press, Taylor
and Francis Group, Boca Raton, FL (2010). ISBN 978-1420085921
25. Jiang, H., Li, X.: Discriminative learning in sequential pattern recognition—a unifying re-
view for optimization-oriented speech recognition. IEEE Signal Process. Mag. 27(3), 115–127
(2010)
26. Jiang, H., Li, X., Liu, C.: Large margin hidden markov models for speech recognition. IEEE
Trans. Audio, Speech Lang. Process. 14(5), 1584–1595 (2006)
27. Juang, B.H., Levinson, S.E., Sondhi, M.M.: Maximum likelihood estimation for mixture mul-
tivariate stochastic observations of markov chains. In: IEEE International Symposium on In-
formation Theory vol. 32(2), pp. 307–309 (1986)
28. Kenny, P.: Joint factor analysis of speaker and session variability: theory and algorithms. CRIM,
Montreal, (Report) CRIM-06/08-13 (2005)
29. King, S., Frankel, J., Livescu, K., McDermott, E., Richmond, K., Wester, M.: Speech production
knowledge in automatic speech recognition. J. Acoust. Soc. Am. 121, 723–742 (2007)
30. Lee, L.J., Fieguth, P., Deng, L.: A functional articulatory dynamic model for speech produc-
tion. In: Proceedings of International Conference on Acoustics, Speech and Signal Processing
(ICASSP), vol. 2, pp. 797–800. Salt Lake City (2001)
31. Rasmussen, C.E.: The infinite gaussian mixture model. In: Proceedings of Neural Information
Processing Systems (NIPS) (1999)
32. Reynolds, D., Rose, R.: Robust text-independent speaker identification using gaussian mixture
speaker models. IEEE Trans. Speech Audio Process. 3(1), 72–83 (1995)
33. Xiao, L., Deng, L.: A geometric perspective of large-margin training of Gaussian models. IEEE
Signal Process. Mag. 27, 118–123 (2010)
34. Yin, S.C., Rose, R., Kenny, P.: A joint factor analysis approach to progressive model adaptation
in text-independent speaker verification. IEEE Trans. Audio Speech Lang. Process. 15(7),
1999–2010 (2007)
http://www.springer.com/978-1-4471-5778-6