Li S.Z., Jain A.K. (eds.) Encyclopedia of Biometrics

Подождите немного. Документ загружается.

France, and China. Other participants have been from

Canada, various European countries, Singapore,

Australia, Israel, and South Africa.

Most of the sites participating in the NIST evalua-

tions have been from academic institutions. Some

government funded research institutions or companies

involved in government research have also partici-

pated. Not frequently represented, however, have

been smalle r commercially oriented companies. This

may be due in part to the text-independent and re-

search oriented type of evaluation being conducted,

but also bespeaks a reticence to participate in evalua-

tions where competitors may show superior perfor-

mance results.

Evaluation requires a performance measure. For

detection tasks there are inherently two types of error.

There are trials where the target is present (target

trials) but a ‘‘false’’ decision is made by a system.

Such errors are misses. And there are trials where the

target is not present (non-target or impostor trials) but

a ‘‘true’’ decision is made. These are referred to as false

alarms. Thus it is possible to speak of a miss rate for

target trials and a false alarm rate for non-target trials.

The NIST evaluations have used a linear combi-

nation of these two error rates as its primary evaluation

metric. A decision cost function (DCF) is deﬁned as

DCF ¼ C

Miss

 P

Miss Targetj

 P

Target

þ C

FalseAlarm

 P

FalseAlarm NonTarget

ð1  P

Target

where C

Miss

represents the cost of a miss, C

the cost

of a false alarm, and P

Target

the prior probability of a

target trial. These are three somewhat arbitrary and

certainly application dependent parameters. The NIST

evaluations have used parameter values as hereunder.

These have been viewed as reasonable parameters

for applications involving an unaware user, where

most spe ech segments examined are likely to be of

someone other than the target of interest, but where

detecting instances of the target have considerable

value. Note that P

Target

need not represent the actual

target richness of the evaluation trials, but may be

chosen based on possible applications of interest. The

NIST evaluations have generally had an approximately

ten to one ratio of non-target to target trials, to mini-

mize the variance of the metric in the light of the

parameter values chosen.

A detection task inherently involves two types of

error, and a system may be expected to be able to tune

its performance to vary the relative frequency of the

two error types. In the NIST evaluations, systems have

been required to produce not only a decision, but also

a score for each trial, where higher scores indicate

greater likelihood that the correct decision is ‘‘true’’.

A decision threshold may then be varied based on this

score to show different possible operating points or

tradeoffs between the two types of error. Note that the

evaluations have required that this threshold be the

same for all target speakers.

The most informative way of presenting system

performance in the NIST SRE’s has been to draw a

curve showing the operating points and the tradeoff in

the error rates. This is easily done by varying the

decision threshold based on the scores provided. A

simple linear plot is know as an ROC (Receiver Opera-

tor Characteristic) curve, but a clearer presentation is

obtained by putti ng both error rates on a normal

deviate scale to produce what NIST has denoted a

DET (Detection Error Tradeoff) curve [4]. This has

the nice property that if the underlying error distri bu-

tions for the miss and false alarm rates are normal, the

resulting curve is linear.

Figure 1 shows

▶ DET curves for the systems in the

core test done in the 2006 NIST SRE. These a re curves

representing the performance of the primary systems

submitted by over 30 sites participating in the evalua-

tion. Better systems have performance curves closer

to the lower left corner of the plot. The actual decision

point of each performance curve is denoted by a

triangle, and a 95% conﬁdence box is drawn around

these, while circles are used to denote the points

corresponding to the minimum DCF operating points.

The closer these two specially denoted points on each

curve, the better the system did at calibrating its deci-

sion threshold for hard decisions. For example, for the

best performing system shown, the actual decision

point has a false alarm rate of about 2% and a miss

rate of about 7%, while the minimum DCF point has a

false alarm rate of about 1% and a miss rate of about

11%. This gives a sense of the level of current state-of-

the-art performance for speaker detection on this type

of telephone data.

Miss

FalseAlarm

Target

10 1 0.01

1250

Speaker Databases and Evaluation

A possible alternative non-parametric information

theoretic type of metric has been proposed to be appli-

cable to a range of applications, and has been included

as an alternative measure in the most recent NIST

evaluations, provided the system speciﬁes that its like-

lihood scores may be viewed as log likelihood ratios.

This metric is discussed in [5].

While the basic detection task has remained ﬁxed,

there have been multiple test conditions in most

of the evaluations, and these conditions have varied

over the years. In particular there has been variation

in the durations of the training and test segments.

While the earlier evaluations focused on landline

phones and the varying types of telephone handsets

(carbon-button vs. electrets microphone), in the new

millennium there is was greater focus on the effect of

cellular transmission and newer types of handsets as

these became common in the U.S. Certain additional

data sources, such as a small FBI forensic database and

a Castilian Spanish corpus know n as AHUMADA

(both apparently not currently easily available) were

used in one or two evaluations.

The earlier evaluations used ﬁxed durations of

speech, as determined by an automatic speech detector.

Later evaluations allowed more variation in duration

within each test condition. Starting in 2001 there was

greater interest in longer durations for training and

test. This was largely as a result of some research

suggesting that with effective word recognition, higher

level lexical information about a speaker could be

effectively combined with more traditional lower level

acoustic information [6]. As a result of the apparent

success of such an approach in the 2001 evaluation,

a major summer research program was carried out at

Johns Hopkins University in the summer of 2001 (see

http://www.clsp.jhu.edu/ws2002/groups/supersid/).

Since then, ‘‘extended’’ training conditions, where the

training consists of multiple (often eight) conversation

sets have been a major part of the evaluations. The

earlier NISTevaluations are described further in [7–9].

The introduction of Mixer data in 2004 inaugu-

rated a new era in the NIST evaluations. The inclusion

of calls in multiple languages and cross language

trials introduced a new wrinkle that affected overall

Speaker Databases and Evaluation. Figure 1 DET (Detection Error Tradeoff) Curves for the primary systems of

participating sites on the core test of the 2006 NIST SRE.

Speaker Databases and Evaluation

1251

performance. The latest evaluations have also intro-

duced test conditions involving multiple microphones

and cross channel trials, that will be a major focus

in 2008 and beyond. The recent SRE’s are discussed

in [10–12].

Have the evaluations shown progress in perfor-

mance capabilities over the years? They have, but

changes in the test conditions from year to year and

in the types of data used have complicated perfor-

mance comparisons. Figure 2 from [13] attempts to

sort these matters out, and summarizes the DCF scores

of the best evaluation systems across ranges of years

involving more or less consistent test conditions.

The NIST SRE’s have been the most notable eva-

luations in speaker recognition in recent years. They

have concentrated on a basic speaker detec tion task

not tied to any speciﬁc current commercial applica-

tion. This has made it possible for a large range of

research sites around the world to participate in these

evaluations.

One other notable evaluation in the ﬁeld was con-

ducted by TNO in the Netherlands in 2003. It featured

a protocol ver y similar to that of the NIST evaluations,

but utilized actual forensic data provided by the Dutch

police. Its very interesting results are discussed in [14],

but the data used was only provided to the evaluation

participants for a limited time and purpose and is not

otherwise available.

Other efforts have been less successful. Research in

speaker recognition technology has been advanced by

the series of Odyssey workshops. These were held

in Martigny, Switzerland in 1996, Avignon, France in

1998, Crete, Greece in 2001 (where the name ‘‘Odys-

sey’’ was adopted), Toledo, Spain in 2004, San Juan,

Puerto Rico in 2006, and Stellenbosch, South Africa

in 2008. For the 2001 workshop an evaluation track

was include d. This included both a text-independent

track based on the preceding NIST evaluation, and

a text-dependent track. Participation, particularly in

the text-dependent track, was very limited, perhaps

demonstrating the difﬁculty of persuading companies

or organizations to participate in this inherently appli-

cation speciﬁc and more immediately commercially

oriented ﬁeld.

The European Union has sponsored a multi-year

program to develop biometric technologies deno-

ted as ‘‘BioSecure’’ (http://www.biosecure.info/), with

speaker as one of the included technologies. Evaluation

is intended to be part of this program, in particular

including evaluation of the fusion of multiple bio-

metrics. As of 2007, however, speaker recognition

evaluation appears not to have begun.

Speaker Databases and Evaluation. Figure 2 DCF (Decision Cost Function) values for the best (lowest DCF) systems

on different roughly comparable evaluation conditions over multiple years during the course of the NIST SRE’s from

1996 to 2006.

1252

Speaker Databases and Evaluation

The NIST evaluations will resume in 2008, and

may be held in alternate years in the future. They will

feature an increased emphasis on cross channel recog-

nition. Whereas in 2005 and 2006 the core test involved

only telephone speech, with cross channel (tr ain on

telephone, test on microphone) as an optional addi-

tional test, the core test condition is expected to

require processing of a mix of training or test segments

including both telephone and microphone speech,

with some of the trials including different channels in

training and test. This will utilize at least both types of

data as in Mixer 3 and Mixer 5. Evaluation perfor-

mance, however, will be subsequently analyzed to dis-

tinguish performance on telephone, microphone, and

cross-channel trials. A number of different micro-

phone types from the Mixer 5 data will be included.

Related Entries

▶ Performance Evaluation, Overview

▶ Speaker Recognition, Overview

References

1. Cieri, C., Campbell, J.P., Nakasone, H., Miller, D., Walker, K.:

The Mixer Corpus of Multilingual, Multichannel Speaker

Recognition Data, LREC 2004: Fourth International Con-

ference on Language Resources and Evaluation, Lisbon (2004)

2. Cieri, C., Andrews, W., Campbell, J.P., Doddington, G., Godfrey,

J., Huang, S., Liberman, M., Martin, A., Nakasone, H.,

Przybocki, M., Walker, K.: The Mixer and Transcript Reading

Corpora: Resources for Multilingual, Crosschannel Speaker Rec-

ognition Research, LREC 2006: Fifth International Conference

on Language Resources and Evaluation (2006)

3. Cieri, C., Corson, L., Graff, D., Walker, K.: Resources for New

Research Directions in Speaker Recognition: The Mixer 3, 4 and

5 Corpora, Interspeech 2007, Antwerp (August 2007)

4. Martin, A.F., et al.: The DET cur ve in assessment of detection

task performance. In: Proceedings of Eurospeech ’97, vol. 4,

pp. 1899–1903. Rhodes, Greece (September 1997)

5. Brummer, N., du Preez, J.: Application-independent evaluation

of speaker detection. Comput. Speech Lang. 20(2–3), 230–275

(April–July 2006)

6. Doddington, G.: Speaker recognition based on idiolectal differ-

ences between speakers. In: Proceedings of Eurospeech ’01,

vol. 4, pp. 2521–2524. Aalborg, Denmark (September 2001)

7. Martin, A.F., Przybocki, M.A.: The NIST speaker recognition

evaluations: 1996–2001. In: Proceedings of 2001: A Speaker

Odyssey, pp. 39–43. pp. 39–43. Chainia, Crete, Greece ( June

2001)

8. Martin, A.F., Przybocki, M.A., Campbell, J.P.: The NIST speaker

recognition evaluation program. In: Wayman, J. (eds.) et al.:

Biometric Systems: Technology, Design and Performance Evalu-

ation, Chapter 8, pp. 241–262. ?pp. 241–262. Springer, Berlin

(2005)

9. Przybocki, M.A., Martin, A.F.: NIST speaker recognition evalua-

tion chronicles. In: Proceedings of Odyssey 2004: The Speaker

and Language Recognition Workshop. Toledo, Spain (2004)

10. Przybocki, M.A., Martin, A.F., Le, A.N.: NIST speaker recogni-

tion evaluation chronicles – Part 2. In: Proceedings of Odyssey

2006: The Speaker and Language Recognition Workshop.

San Juan, PR (2006)

11. Przybocki, M.A., Martin, A.F., Le, A.N.: NIST speaker recogni-

tion evaluations utililizing the mixer corpora – 2004, 2005, 2006.

IEEE Trans. Audio Speech Lang. Process. 15(7), (2007)

12. Martin, A.F.: Evaluations of automatic speaker classiﬁcation

systems. In: Muller, C. (ed.) Speaker Classiﬁcation I, pp. 313–

329. pp. 313–329. Springer, Berlin (2007)

13. Reynolds, D.A.: Keynote talk. In: Proceedings of Odyssey

2008: The Speaker and Language Recognition Workshop.

Stellenbosch, South Africa (January 2008)

14. van Leeuwen, D.A., et al.: NIST and NFI-TNO evaluations

of automatic speaker recognition. Comput. Speech Lang.

20(2), 128–158 (2006)

Speaker Detection

Speaker detection means determining whether or not a

particular speaker is present in an audio stream. The

term multispeaker detection refers to the task of deter-

mining whether a particular known speaker is speaking

in an audio stream containing speech from multiple

speakers.

▶ Speaker Segmentation

Speaker Diarization

This task consists of segmenting a conversation involv-

ing multiple speakers into homogeneous parts, which

contain the voice of only one speaker, and grouping

together all the segments that correspond to the same

speaker.

▶ Speaker Segmentation

Speaker Diarization

1253

Speaker Features

DANIEL RAMOS,JAV I E R GONZALEZ-DOMINGUEZ,

OROTEO T. TOLEDANO,

OAQU I N GONZALEZ-RODRI

GUEZ

ATVS – Biometric Recognition Group. Escuela

Politecnica Superior, Universidad Autonoma de

Madrid, Spain

Synonyms

Observations from speech; Speaker parameters

Definition

Speaker features are measurements extracted from the

speech signal with the objective of determi ning the

identity of a given speaker. In voice biometrics, speaker

features whose source is known are typically used to

build

▶ speaker models. Then, speaker features of un-

known source are compared with the enrolled models

in order to obtain measures of similarity. The identity

of the speaker inﬂuences the speech production pro-

cess in many different ways, due to vocal tract conﬁgu-

ration, language spoken, social context, education, etc.

Thus, several levels of identity can be identiﬁed in the

speech signal, e.g., spectral, phonetic, prosodic, etc.

Speaker features can be extracted at any of this identity

levels, and therefore the speaker recognition process

follows in essence a multilevel approach.

Identity Information in the Speech

Signal

The ▶ identity levels in the speech signal are conﬁg-

ured by the speech production process, which is the

subject of study of phoneticians and other areas such as

engineering, physics or signal processing [1, 2, 3, 4].

There are two main stages in voice production: (1) lan-

guage generati on and (2) speech production; and

speaker speciﬁ cities are introduced in both compo-

nents. In the ﬁeld of speaker recognition these two

components correspond to so-called high-level (lin-

guistic) and low-level (spectral) characteristics. Auto-

matic speaker recognition systems will intend to take

advantage of the different sources of information avail-

able in the speech signal, combining them in the best

possible way for every speaker [5, 6]. Figure 1 illus-

trates these different identity levels in the speech signal.

The information extracted in each of these groups of

levels can be summarized as follows:



Spectral level. The information about the speaker

identity is extracted from the spectrum of the

speech signal, analyzed in short-time windows.

The spectrum of the speech signal is directly related

to the dynamic conﬁguration of the vocal tract,

which presents speaker-dependent speciﬁcities.



Higher levels. Several sublevels can be found

here. For instance, at the phonotactic level, the infor-

mation about the identity of the speaker is embedded

in the particular use of the phones and syllables and

their realizations. At the prosodic level, parameters

like instantaneous energy, intonation, speech rate,

and unit durations are analyzed, which are known

to be speaker-dependent. At the idiolectal level, the

information about speaker identity relies on the par-

ticular use of the words and language in general,

which depends not only on the speaker, but also on

many other sociolinguistic conditions.

Short-Term Spectral Feature

Extraction

The analysis at spectral level of the speech signal is

based on classic Fourier analysis. However, an exact

deﬁnition of Fourier transform cannot be directly ap-

plied becaus e speech signal cannot be considered

Speaker Features. Figure 1 Identity levels in the speech

signal.

1254

Speaker Features

stationary due to constant changes in the articulatory

system within each speech utterance.

To solve these problems, speech signal is split into a

sequence of short segme nts in such a way that each one

is short enough to be considered pseudo-stationary.

The length of each segment, also called window or

frame, ranges between 10 and 40ms (in such a short-

time period our articulator y system is not able to

signiﬁcantly change). Finally, a feature vector will be

extracted from the shor t-time spectrum in each win-

dow. The whole process, known as short-term analysis,

is depicted in Fig. 2.

Signal representation or coding from short-term

spectrum into a feature vector is one of the most

important steps in automatic speaker recognition and

continues being subject of research. Many different

techniques have be en proposed in the literature and

generally they are based on speech production models

or speech perception models. Most widely used tech-

niques in the state of the art are described as follows.



Linear Predictive Coding (LPC) method, introduced

in [7], is based on the assumption that a speech

sample can be approximated by a linearly weig hted

summation of a determined number of preceding

samples. In time domain, this can be represented as



½n¼

k¼0

ak½sn k½: ð1Þ

Here, s

∗

[n] is the approximation, or prediction,

of the speech signal, and a [k] are the LPC coefﬁ-

cients calculated to minimize the total square error

E ¼

e½n

; ð2Þ

where e[n] is the error between the real signal

value s [n] and predicted value s

∗

[n], deﬁned as

e½n¼s½ns



½n¼s½n

k¼1

ak½sn k½: ð3Þ

In the domain of the z-transform, a [k] para-

meters deﬁne an all-pole ﬁlter H(z), as deﬁned

in [1, 7].

HzðÞ¼

1 

k¼1

ak½z

k

: ð4Þ

LPC has proved to be a valid way to compress

the spectral envelope in an all-pole model with just

10–16 coefﬁcients [1, 3]. Figure 3 shows the rep-

resentation of the envelope of the short-time spec-

trum at a given window as modeled by LPC.

However, LPC coefﬁcients are strongly correlated

among them, which is an undesirable characteris-

tic. Therefore,

▶ cepstrum transform [3, 8] has

been proposed in order to obtain pseudo-orthogo-

nal cepstral coefﬁcients, yielding Linear Prediction

Cepstral Coefﬁcients (LPCC).



Mel-Frequency Cepstral Coefﬁcients (MFCC) pro-

posed in [9] are the most extensively used para-

meters at the spectral level in automatic speaker

recognition systems. The MFCC method ﬁrst uses a

mel-scale ﬁlterbank in order to obtain some coefﬁ-

cients from the power spectrum of the speech win-

dow. The main aim of mel ﬁltering is to mimic

Speaker Features. Figure 2 Short-term feature extraction.

Speaker Features

1255

the human hearing behavior by emphasizing lower

frequencies and penalizing higher frequencies. Thus,

a mel ﬁlterbank analyzes the power spectrum using a

logarithmic scale. First, a transformation is applied

according to the following formula:

¼ 1; 125  log 1 þ f =700ðÞ; ð5Þ

where f is the linear frequency . Second, a ﬁlterbank is

applied to the amplitude of the mel-scaled spectrum f

order to obtain a vector of outputs from each ﬁlter .

Figure 4 shows a typical mel ﬁlterbank in the fre-

quency domain. The centers f [m] of the ﬁlters H

[k]

are uniformly spaced in the mel scale. Using a DFT

of the input signal withN points each ﬁlter H

[k]is

given by

k½¼

0 k < fm1½

kfm1½ðÞ

fm½fm1½ðÞ

fm1½k fm½

fmþ1½kðÞ

fmþ1½fm½ðÞ

fm½k fm½

0 k > fmþ1½;

ð6Þ

where 0<k<N.

Once ﬁltering is carried out, cepstrum transform is

applied to the ﬁlter outputs in order to obtain mel

frequency cesptrum coefﬁcients.

Speaker Features. Figure 3 LPC modeling of short-term spectrum. The envelope is determined by the LPC filter

coefficients defined in Eq. 4 .

Speaker Features. Figure 4 Triangular mel-filter bank for typical MFCC feature extraction.

1256

Speaker Features



Perceptual Linear Prediction (PLP) was proposed

in [10]. Here, speaker features are calculated in a

similar way as LPC coefﬁcients, but previous trans-

formations are carried out in the spectrum of each

window aiming at introducing knowledge about

human hearing behavior. Details can be found in

[10].

As we mentioned earlier, the main aim of the

described methods is to extract a feature vector for

each frame or window. However, in this independent

analysis possible useful information such as coarticula-

tion can be lost. In order to take this kind of information

into account, velocity (D) and acceleration (DD) coefﬁ-

cients are usually obtained from the static window-

based informati on. This D and DD coefﬁcients model

the speed and acceleration of the variation of cepstral

feature vectors across adjacent windows.

High-Level Tokenization

At phonotactic and idiolectal levels, tokenization is the

translation from sampled speech into a time-aligned

sequence of linguistic units, or tokens. Hidden Markov

Models (HMM) [11] are widely used for phone, sylla-

ble, and word tokenization. HMM as used in speech

processing are ﬁnite stat e machines which model the

temporal dependency of spectral feature vectors in a

probabilistic way [1, 11 ]. The performance of the toke-

nization may be improved by the use of language

models, which impose some linguistic or grammatical

constraints on the high number of combinations of all

possible units (phones, sy llables or words) [1].

Basic prosodic features as pitch and energy are also

obtained at a frame level. The instantaneous pitch

can be determined by, e.g., autocorrelation or cepstral-

decomposition based methods, usually smoothed with

some time ﬁltering [2]. Other important prosodic fea-

tures are those related to linguistic units duration,

speech rate, and all those related to accent. In all

those cases, precise segmentation is required [1, 12],

i.e., determination of the points in the speech signal

where each unit occurs.

Recently, Nonuniform Extraction Region Features

(NERF) have been proposed for obtaining high level

features [13]. This technique is based on including

high-level information in the spectral information at

each short-time frame.

Summary

The information about the identity of the speaker

extracted from a speech utterance is represented by

the speaker features, which can be obtained at different

levels. The essay presented the main approaches for

speaker feature extraction at the short-time spectral

level and at higher levels. The widely used MFCC,

LPCC, and PLP features have been described, and

several approaches of phonetic and prosodic tokeniza-

tion have been sketched. Such features will be used to

build the models and to compare them with test speech

segments in a given voice biometric system.

Related Entries

▶ Feature Extraction

▶ Speaker Matching

▶ Session Effects on Speaker Modeling

▶ Speaker Recognition, Overview

▶ Speech Analysis

References

1. Huang, X., Acero, A., Hon, H.W.: Spoken Language Processing:

A Guide to Theory, Algorithm and System Development. Pren-

tice Hall PTR, Upper Saddle River, NJ (2001)

2. Rabiner, L.R., Schafer, R.W.: Digital Processing of Speech Sig-

nals. Prentice Hall, Englewood Cliffs, NJ (1978)

3. Deller, J.R., Hansen, J.H.L., Proakis, J.L., Discrete-Time Proces-

sing of Speech Signals, 2nd Ed. Wiley, New York (1999)

4. Gonzalez-Rodriguez, J., Toledano, D.T., Ortega-Garcia, J.: Voice

biometrics. In: Anil K. Jain, Patrick Flynn and Arun A. Ross

(eds.) Handbook of Biometrics. Springer, Berlin (2007)

5. Reynolds, D.A.: The SuperSID project: Exploiting high-level

information for high-accuracy speaker recognition. In: Proceed-

ings of ICASSP (2003)

6. Doddington, G.: Speaker recognition based on idiolectal differ-

ences between speakers. In: Proceedings of Eurospeech,

pp. 2517–2520. Aalborg, Denmark (2001)

7. Makhoul, J.: Spectral analysis of speech by linear prediction.

IEEE Trans. Audio Electroacoust. 21, 140–148 (1973)

8. Furui, S.: Cepstral analysis technique for automatic speaker

veriﬁcation. IEEE Trans. Acoust. Speech, Signal Processing 29,

254–272 (1981)

9. Bridle, J.S., Brown, M.D.: An experimental automatic word

recognition system. Technical Report 1003, Joint Speech Re-

search Unit, Ruislip, Eng land (1974)

10. Hermansky, H., Hanson, B., Wakita, H.: Perceptually based

linear predictive analysis of speech. In: Proceedings of ICASSP,

Vol.10, pp. 509–512. (1985)

Speaker Features

1257

11. Rabiner, L.R.: A tutorial on hidden markov models and selected

applications in speech recogniton. Proc. IEEE 77, 257–286 (1989)

12. Toledano, D.T., Hernandez-Gomez, L., Villarrubia-Grande, L.:

Automatic phonetic segmentation. IEEE Trans. Speech Audio

Process 11, 617–625 (2003)

13. Kajarekar, S., Ferrer, L., Sonmez, K., Zheng, J., Shriberg, E.,

Stolcke, A.: Modeling NERFs for speaker recognition In: Pro-

ceedings of Odyssey, pp. 51–56 Toledo, Spain (2004)

Speaker Identification and

Verification, SIV

▶ Speaker Recognition, Standardization

Speaker Indexing

The process of determining, who is talking when, is an

integral element of speech data monitoring and

content-based data mining applications.

▶ Speaker Segmentation

Speaker Matching

JEAN-FRANC¸OIS BONASTRE,DRISS MATROUF

LIA - CERI University of Avignon, Avignon, France

Synonyms

Speaker recognition engine; Voice biometric engine

Definition

Speaker matching aims to compare the acquired data

corresponding to an individual against the template

feature set stored in the database. Depending on the

operating mode, the comparison could be done using

only the template related to a given person (detection

or veriﬁcation tasks) or with all the templates of the

database (identiﬁcation task).

The speaker matching could be split into three

main functionalities:



Create a template from the feature set extracted

from the enrollment data. Usually, the template is

denoted ‘‘speaker model.’’



Compare a feature set acquired from a sound cap-

tor with a speaker model and output a likelihood

score.



Take an identiﬁcation decision using this score.

Usually other information are used during this

decision phase, like a model of potential impostors.

In a speaker recognition system, the scores are very

often normalized before to take the decision, using

▶ Score Normalization technique.

Introduction

Two main classes of approaches are usually used for

speaker matching:



Direct matching. This kind of approaches don’t

really use a modeling of the speaker voice: the en-

rollment voice sample is used directly as a model

and a similarity function between two voice samples

is used for the score computation. If the time syn-

chronization aspects are taken into account, a dy-

namic time warping algorithm is used in order to

ﬁnd the best time alignment between the enroll-

ment acoustic feature sequence and the test feature

test sequence.

▶ Dynamic Time Warping (DTW)

[1, 2] involves a strong dependency on the text

pronounced by the individuals: the text should be

the same during the enrollment and the recogni-

tion phases, the message should have a short dura-

tion (few seconds). The Fig. 1 illustrates the DTW

main principles. For text independent systems,

where the text pronounced by the individuals

could be different between the enrollment and the

recognition phases, a

▶ Vector Quantiﬁcation

(VQ) algorithm is used [2]. The main principle of

VQ is illustrated in Fig. 2. In this case, the VQ

codebook could be seen as a model, but it is in

fact close to a data reduction of the enrollment

sample: only a subset of the feature vectors

extracted from the enrollment data are kept and

the other are withdrawn. The main advantage of

these methods is the low level of needed computer

resources. These methods are also efﬁcient in terms

1258

Speaker Identification and Verification, SIV

of identiﬁcation performance if the variability fac-

tors like utterance text content, microphone or

environmental noises are controlled [1, 3].



Machine learning approaches. In this class of meth-

ods, a speaker voice model is learnt using one or

several enrollment recordings. During the test, the

likelihood of the test data is computed using this

model. Two kind of methods are used: the statis-

tical modeling techniques (mainly GMM or HMM)

and the discriminative classiﬁcation techniques

(mainly neural networks and SVM). Recently,

mixed approaches were proposed, where a statistical

approach (based on GMM) is used to deal with the

data variability and a discriminative classiﬁer is used

for the decision estimation (SVM).

This chapter will describe more precisely the

Machine Learning based approaches, which associate

the GMM and the SVM.

GMM-UBM (GMM-MAP) Approach

GMM-UBM is the predominate approach used in

speaker recognition systems, particularly for text-

independent task [4]. This approach is based on a

Speaker Matching. Figure 1 DTW principle.

Speaker Matching. Figure 2 Frames distribution of two speakers in the acoustic space. Each speaker can be

characterized by its own distribution modes (centroids).

Speaker Matching

1259