Frame Selection for Text-independent Speaker Recognition

Abedenebi Rouigueb, Malek Nadil and Abderrahmane Tikourt

Applied Mathematics Laboratory, Ecole Militaire Polytechnique, Algiers, Algeria

Keywords:

Frame Selection, Frame Pruning, Speaker Identiﬁcation, Speaker Recognition, Text-independent.

Abstract:

In this paper, we propose a set of criteria for the selection of the most relevant frames in order to improve

text-independent speaker automatic recognition (TISAR) task. The selection is carried out on the short term

Cepstral feature vectors such as PLP and MFCC and performed at the front end processing level. The proposed

criteria mainly attempt to select vectors lying far from the universal background model (UBM). Experiments

are conducted on the MOBIO database and show that the selection allows an improvement in complexity (time

and space) and in speaker identiﬁcation rate, which is appropriate for real-time TISAR systems.

1 INTRODUCTION

Text-independent automatic speaker recognition

(TIASR) task consists in verifying or in identifying

the speaker identity using a segment of his speech

where the utterance content is free (Beigi, 2011; Kin-

nunen and Li, 2010). Although a wealth of research

works has addressed this problem throughout the last

40 years, it is still a challenging problem with many

potential applications.

Cepstral short-term features extracted from short

time frames of about 20-30 milliseconds in duration,

such as Mel frequency cepstral coefﬁcients (MFCC)

and perceptual linear prediction (PLP) coefﬁcients

(Kinnunen and Li, 2010), are widely used in TIASR

systems. This choice is in part justiﬁed by the non-

stationary nature of voice and the free speech as-

sumption. The order between the feature vectors (e.g

MFCC) is usually not utilized; the TIASR is merely

treated as a recognition problem of a set, and not a se-

quence, of the acoustic vectors coming from the test

utterance.

In typical TIASR systems, voice activity detection

(VAD) (Benyassine et al., 1997) is ﬁrst applied for re-

moving non-speech frames. Unfortunately, due to the

high variability and the non deterministic nature of

noise, VAD cannot be perfectly achieved in all cases.

VAD is more critical for non-stationarynoise environ-

ments. Hence, some signal segments corresponding

to noise and not to speech might be left. For further

details about the challenges facing VAD techniques,

refer to (Ramrez et al., 2004). Moreover, the noise

cannot be completely ﬁltered from the speech signal,

either because the signal is too much altered or be-

cause the noise type is unknown. For useful discus-

sions, see (Loizou, 2013).

In the TIASR context, different phonemes can be

pronounced within a given utterance where each one

produces about 5-10 Cepstral vectors. It has been

shown that nasal consonants and vowels are more

speaker-discriminative than the other phonemes like

stops. Indeed, the speaker discrimination quality de-

pends closely on the phonetic content. A quantitative

assessment of phoneme groups for speaker recogni-

tion is given in (Eatock and Mason, 1994).

To recap, acoustic vectors resulting from i) non-

speech segments, due to VAD failure; ii) severely

corrupted speech segments; or iii) less-discriminative

phonemes, are less relevant than those extracted from

enhanced speech frames of discriminative phoneme

group.

In light of the above-mentioned considerations,

the TIASR task can be perceived as a classiﬁcation

problem of the acoustic vectors set resulting from the

test utterance where they are involved with different

degrees in the ﬁnal decision.

In this paper, we aim at proposing selection meth-

ods of the most speaker discriminative acoustic vec-

tors (frames). It is obvious that front-end frame se-

lection leads to a signiﬁcant reduction of the spatio-

temporal complexity since the recognition process is

achieved based only on a small subset of vectors. Our

concern is focused here on the assessment of the inﬂu-

ence of the amount of the selected data on the speaker

Rouigueb, A., Nadil, M. and Tikourt, A.

Frame Selection for Text-independent Speaker Recognition.

DOI: 10.5220/0006392100510057

In Proceedings of the 14th International Joint Conference on e-Business and Telecommunications (ICETE 2017) - Volume 5: SIGMAP, pages 51-57

ISBN: 978-989-758-260-8

recognition accuracy using several criteria. The per-

formance is measured in terms of the selection task

time and the speaker identiﬁcation rate in relation to

the percentage of the selected data.

The remainder of this paper is organized as fol-

lows. Section 2 presents related works and Section 3

discusses the utility of frame selection in TISAR sys-

tems. Section 4 shows the importance of the univer-

sal backgroundmodel in the state-of-the-artof TIASR

models. In Section 5 we present the proposed criteria,

experimentations and results are discussed in Section

6. We highlight the improvements done in relation to

our prior works in Section 7. Finally, conclusions and

future directions are given in Section 8.

2 RELATED WORKS

In TISAR systems, a test utterance signal is split into

a sequence of short-time frames and a features vector

is extracted from each frame. Data reduction is often

performed; it may be achieved by two complemen-

tary approaches: feature selection or vector (frame)

selection. A large number of works have addressed

the frame selection problem (also referred to as frame

pruning). For instance, silence frame removal and

VAD are usually applied at an early stage to remove

a part of irrelevant frames. In addition to the spatio-

temporal complexity reduction, an interesting frame

pruning technique must not compromise the recog-

nition performance. Beyond to silence removal and

VAD pruning, we present in this section some inter-

esting works found in the literature that have treated

the frame pruning problem using different criteria.

A test utterance is divided into multiple frames

(Besacier and Bonastre, 1998a) or multiple time-

frequency blocks (Besacier and Bonastre, 1998b) and

the ﬁnal identiﬁcation score is computed with a lim-

ited number of the obtained frames (blocks). A dis-

criminant function, estimated for each speaker, is

used to remove frames (or blocks) having a low log

likelihood ratio score of the speaker model against the

speakers’ background model. Authors have reported

that using only a 30% frame pruning can increase

signiﬁcantly the identiﬁcation rate where the exper-

iments are conducted on the TIMIT and NTIMIT cor-

pora.

In (Kinnunen et al., 2006), the number of vectors

of the utterance test is reduced by silence removable

and pre-quantization (PQ) in order to speed up the

identiﬁcation process. Pre-quantization aims to keep

a subset of vectors using different PQ techniques as

random sub-sampling, averaging and decimation, see

(Kinnunen et al., 2006) for more details. McLaughlin

et al. have shown that the application of three sim-

ple PQ methods, prior to GMM (Gaussian Mixture

Model) matching, allows to compressing the test se-

quence by a factor of 20:1 without compromising the

veriﬁcation accuracy (McLaughlin et al., 1999).

Recently, Almaadeed et al. have proposed a real-

time text-independent speaker identiﬁcation system

where the consonant frames are ﬁltered out and the

identiﬁcation is based on the formants extracted from

vowels (Almaadeed et al., 2016).

3 FRAME SELECTION FOR

TISAR SYSTEMS

If we make a deep analysis of the most popular ac-

curate TIASR paradigms: ranging from the GMM-

UBM baseline (Reynolds et al., 2000) to the state-

of-the-art of the speaker veriﬁcation I-vector con-

cept (Dehak et al., 2011) via the GMM supervector

(Campbell et al., 2006) and the joint factor analysis,

one can conclude that:

1. All these models make use of short term acous-

tic vector,particularly MFCC, where the utterance

vectors are seen as a set;

2. Some regions in the feature MFCC space are pe-

nalized more than others. For instance, in the

GMM-UBM baseline system, MFCC vectors with

a high density within the UBM class are disad-

vantaged in the likelihood decision ratio detector

(Reynolds et al., 2000). In the I-vector paradigm,

a GMM supervector, M, is ﬁrst extracted using

the Bayesian MAP adaptation (Campbell et al.,

2006). Then, an I-vector of low dimension, w,

is computed such that M = m+ Tw where T is a

N × P rectangular mapping matrix and m is the

GMM supervectors mean. There is an alignment

between the entries of M and the mixture com-

ponents centers of the GMM-UBM model. The

signiﬁcant dimension reduction (N << P) is fea-

sible because several entries of the vector M − m

are either zeros, or correlated.

So MFCC vectors with a high likelihood within

the GMM-UBM componentscorrespondingto the

less-discriminative entries of M are penalized in

the decision function.

The frame selection can be interesting for TIASR

systems when the utterances duration is not very

short. In such situation, we need efﬁcient algorithms

requiring low complexity in time and in memory

space. For example, TIASR running on smartphones

may be a typical case of use.

SIGMAP 2017 - 14th International Conference on Signal Processing and Multimedia Applications

From a geometric point of view, the feature MFCC

regions don’t hold the same quantity of the speaker-

dependent information. ALL accurate TIASR sys-

tems based on MFCC attempt, either in an explicit

or an implicit way, to determine the importance of

MFCC vectors according to their location. There-

fore, for an absolute research purpose, proposing new

selection criteria can be useful in order to develop

new features and models by the segmentation of the

MFCC space into regions according to their rele-

vance.

4 UNIVERSAL BACKGROUND

MODEL UTILITY

The concept of the universal background model is

successfully used in (Reynolds et al., 2000), in which

the veriﬁcation task is seen as a test between the hy-

pothesis H0: X is uttered by the claimed speaker

against the alternative hypothesis H1: X is uttered by

an impostor. A GMM, commonly referred as GMM-

UBM model, is trained from a collection of data from

a large number of expected speakers and it is used to

ﬁt the UBM density distribution. Few years later, the

GMM-UBM model has been very successful when

used in representing the speaker-independent infor-

mation rather than the alternative hypothesis H1. The

key idea consists in mapping a given utterance to a

ﬁxed-size GMM supervector.

The GMM-supervector of an utterance is derived

via the MAP adaptation of the GMM-UBM distribu-

tion (Campbell et al., 2006). It has been shown that

the best overall performance is from adapting only the

mixture components means (centers) (Reynolds et al.,

2000) compared to weights and covariance matrices.

First, a GMM λ ﬁtting the distribution of the utterance

MFCC vectors is estimated via the MAP adaptation.

Then, a GMM-supervectoris obtained by the concate-

nation of the mixture componentsmeans of λ. Indeed,

a GMM-supervector deﬁnes the overall location of the

components of λ in relation to the GMM-UBM refer-

ence model.

A GMM-supervector, M, is decomposed into two

components as follows M = m + Tw in the i-vector

model case. m itself is a GMM-supervector cor-

responding to the UBM class; it is discarded in

the recognition process since it represents speaker-

independent information.

GMM-UBM is used to make a mapping of an ut-

terance X to a supervector such as GMM supervector,

JFA factors, or i-vector. The location of the MFCC

vectors of an utterance in relation to the UBM is so

important in these new models. It is worth recalling

that the angle between two supervectors is a relevant

feature; many scoring functions are based on the co-

sine kernel (Dehak et al., 2011)

In this paper, we put forward the proposals

that high density regions in the UBM class con-

tain often MFCC vectors coming from several speak-

ers. Hence, these regions are less discrimina-

tive since they tend to model common information

as: non-speech segments, noisy speech, and non-

discriminative phonemes. In this sense, we suggest to

propose selection criteria by adopting the second in-

terpretation which consists in using UBM to estimate

the importance of MFCC regions.

5 PROPOSED CRITERIA

Let X = {x

,...,x

} be a set of N acoustic vectors ex-

tracted from a given utterance U, where each x

is a

D-dimensional cepstral vector (e.g MFCC, PLP ). For

the sake of simplicity, we propose to select a percent-

age of the vectors of X that maximize the proposed

criteria because they have not the same duration in

general. The proposed criteria are as follows.

1. Standard deviation (C

)

For a feature vector x= (x

,...,x

)

measures

the dispersion of the different entries of x , it is

calculated as follows:

(x) =

∑

i=1

−

, (1)

where

x is the mean value of x entries (x =

i=1

/D).

2. Euclidean distance (C

)

measures the Euclidean distance between the

vector of interest xand the center of UBM, µ.

(x) = kx,µk =

i=1

− µ

)

(2)

3. Probabilistic distance (C

)

A GMM, λ

ubm

, with K mixture components is

trained to approximate the UBM distribution.

(x) is inversely proportional to the likelihood

of UBM. We suggest:

(x) = −P

ubm

(x) (3)

We have experimented the variants C

(1)

, C

(4)

(32)

, C

(64)

, C

(128)

, C

(256)

by setting K to 1, 4, 32,

64, 128, and 256, respectively, in λ

ubm

Frame Selection for Text-independent Speaker Recognition

4. Probabilistic distance with component selection

)

The time cost of the evaluation of C

linearly in-

creases with the number of components in λ

ubm

K, because we need to compute the density of x

in each one of them.

Each component of λ

ubm

corresponds to a speciﬁc

sound or phoneme group. Obviously, some com-

ponents better describe speaker-dependent infor-

mation than others. C

constitutes an improve-

ment of C

, where the key idea consists in com-

posing a new GMM, λ

′

ubm

, only from the less-

discriminative components of λ

ubm

. Hence, we

gain in selection speed and possibly in accuracy.

Let λ

ubm

= {w

,Σ

}

i=1..K

be a K-component

GMM-UBM which is estimated from the speech

of S speakers. First, we compute a confusion ma-

trix H where each cell H(s,k) represents the por-

tion of the likelihood of s which is expressed by

the component k. Formally, we have:

H(s,k) =

P(N(m

,Σ

);Xs)

∑

i=1

P(N(m

,Σ

);Xs)

, (4)

Xs is the data of speaker s. The sum of each row

of H is equal to 1. The standard deviation of the

column of index k measures the variability of the

component k. Low standard deviation means that

the speakers’ likelihood values are close to each

other and that the component k is consequently

less-discriminative.

In experiments, we have built λ

′

ubm

from a quar-

ter of the less-discriminative components of λ

ubm

We have experimented the variants C

(4)

, C

(32)

(64)

, C

(128)

by setting K to 4, 32, 64, 128, re-

spectively. For example, in C

(32)

, λ

′

ubm

contains

only the 08 less-discriminative components of the

32 components forming λ

ubm

. Thus, we suggest:

(x) = −P

′

ubm

(x) (5)

6 EXPERIMENTS

6.1 Dataset

All of the results we report are on the MOBIO

database (McCool et al., 2012). Following the same

spirit as the NIST SRE, the Biometric Group at the

Idiap Research Institute organized the evaluation on

text-independent speaker recognition.

MOBIO is a bimodal (audio and video) database

recorded from 152 persons, 100 males and 52 females

with both native and non-nativeEnglish speakers. For

each individual, 12 sessions were captured where 192

utterances are recorded by mobile phone (NOKIA

N93i). MOBIO is a challenging database since the

data is acquired on mobile devices possibly with real

noise, and the speech segments can be very short (less

than 02 sec ). The average speech duration of MO-

BIO phrases is around 08 sec. MOBIO is designed

so that it contains realistic and common environmen-

tal variations associated with the usage of mobile de-

vices. More details on this dataset could be found in

(Khoury et al., 2013).

6.2 Methodology

We propose to evaluate the performance of the pro-

posed criteria inside a speaker identiﬁcation system.

It is clear that useful information for the speaker iden-

tiﬁcation task, is also useful for the speaker veriﬁca-

tion task and vice-versa.

The vectors selection module is integrated in the

identiﬁcation system as illustrated in Figure 1. A per-

centage of cepstral vectors of each training or test ut-

terance that maximize the selection criterion are used

for identiﬁcation. The selection module may use the

parameters of the GMM-UBM model.

Cepstral

extrcation

Train classifier

(SVM, KNN,etc.)

Cepstral vectors set

(MFCC, PLP,etc.)

Selected vectors

Cepstral vectors set

(MFCC, PLP,etc.)

Selected vectors

Selection

Classification

model

Vectors

classification

Majority vote

Decision

U is uttered by S

Vectors classes

UBM GMM

Test utterance,U

Speakers' train utterances

Cepstral

extrcation

Training (offline)

Testing (online)

Selection

Figure 1: Cepstral data selection for speaker identiﬁcation.

The objective of the experimentation is to study the

identiﬁcation performance as function of the follow-

ing parameters.

1. The percentage of selected data, θ ∈ [1,..,100];

2. The selection criterion,C ∈ {C

, variants ofC

and C

}

SIGMAP 2017 - 14th International Conference on Signal Processing and Multimedia Applications

3. The classiﬁer used, Cl ∈ { SVM , KNN }.

For a given combination (θ, C, and Cl), an iden-

tiﬁcation experiment is carried out using 20 speakers

where 10 and 30 utterances are used in test and train-

ing (resp.) for each speaker. Speakers and utterances

involved in experiments are drawn randomly from the

whole MOBIO corpus. Then, the following perfor-

mance measures are computed:

• The identiﬁcation rate :

τ =

number o f correct identification trials

number o f alltrials = 20∗ 30 = 600

;

• The time required for vectors selection of the test

utterances, T

sel

;

• The time required for vectors classiﬁcation of the

test utterances, T

6.3 Experimental Results & Discussions

In this section, we present experiments and results ob-

tained in order to observe the inﬂuence of the selec-

tion parameters over the identiﬁcation performance.

Each identiﬁcation experiment is conducted as shown

in subsection 6.2.

The ﬁrst set of experiments examine the variation

of the identiﬁcation rate, τ, as function of the selection

criterion, C, and percentage, θ. Used features consist

of 19 MFCCs acoustic vector, computed with a frame

shift of 10 ms and a frame size of 25 ms. The identiﬁ-

cation rate curves when SVM (resp. KNN) is applied

are depicted in Figure 2 (resp. 3).

θ : data percentage

0 10 20 30 40 50 60 70 80 90 100

τ : identiﬁcation rate

0.3

0.4

0.5

0.6

0.7

0.8

0.9

128

256

128

Figure 2: Accuracy as function of criteria and data percent-

age using SVM classiﬁer.

While the difference between criteria in terms of τ

for both ﬁgures is not large, C

variants performed

slightly better than the rest. This seems logical since

attempts to keep only vectors having low likeli-

hood in the less-discriminative mixture components

of the GMM-UBM model.

The high rates obtained by C

and C

are surpris-

ing and prove that the GMM-UBM distribution has

roughly a hyper-spherical form which is much dense

around the center.

θ : data percentage

0 10 20 30 40 50 60 70 80 90 100

τ : identiﬁcation rate

0.3

0.4

0.5

0.6

0.7

0.8

0.9

128

256

128

Figure 3: Accuracy as function of criteria and data percent-

age using KNN classiﬁer.

θ data percentage

0 10 20 30 40 50 60 70 80 90 100

Classiﬁcation time (sec)

1000

2000

3000

4000

5000

6000

7000

8000

128

256

128

Figure 4: Classiﬁcation time, T

,as function of data per-

centage, θ.

θ data percentage

0 10 20 30 40 50 60 70 80 90 100

Selection time (sec)

128

256

128

Figure 5: Selection time, T

sel

,as function of data percentage,

θ.

When examining the variations of τ of all criteria in

both ﬁgures 2 and 3, one can clearly see three distinct

regions of θ, as follows:

a) θ ∈ [0% − 5%]: the identiﬁcation rates obtained

are too poor because the selected data are not suf-

ﬁcient in amount for the recognition task;

b) θ ∈ [5%− 25%]: the highest rates are achieved in

this interval;

c) θ ∈ [25%−100%]: a continuous deterioration of τ

is seen by the increasing of θ value. Acoustic vec-

tors minimizing the proposed criteria don’t hold

enough of speaker-dependent information, their

incorporation in the recognition task leads to a de-

creasing in performance.

In Figure 4, we show the classiﬁcation time (for

Frame Selection for Text-independent Speaker Recognition

600 identiﬁcation test trials) using SVM. The time

curves are quite identical for all criteria and seem to

have a quadratic augmentation. Indeed, the classiﬁca-

tion time depends mainly on the data percentage, θ.

In Figure 5, we show the selection time. As expected,

the fastest criteria are C

, C

because few op-

erations are needed to evaluate them. The test time in-

cludes principally selection and classiﬁcation. There-

fore, for a global comparison, C

, represent a

good trade-off between accuracy and complexity. We

deduce that taking θ in [10% 25%] is interesting in the

MOBIO corpus case.

Table 1: Best identiﬁcation rates using PLP, MFCC+∆ fea-

tures.

PLP MFCC+∆

SVM KNN SVM KNN

τ 0.9550 0.8900 0.9775 0.8550

∗

,θ

∗

,3% C

256

,18% C

256

,15%

Table 2: The inﬂuence of the intersession variability.(C =

, classiﬁer=SVM).

sessions τ

∗

,θ

∗

sessions τ

∗

,θ

∗

1 0.990, 07% 7 0.990, 10%

2 0.990, 20% 8 0.990, 07%

3 0.995, 04% 9 1.000, 04%

4 0.990, 12% 10 0.990, 10%

5 1.000, 07% 11 1.000, 16%

6 0.990, 14% 12 1.000, 07%

all(1-12) 0.95, 16%

The same set of experiments is achieved on the PLP,

MFCC+delta features. We have noticed the same be-

havior for both cases where the best results are ob-

tained by setting θ ≤ 20%, see Table 1.

The last set of experiments aims to explore the in-

ﬂuence of selection towards the session differences.

The intersession variability is well-known to be a

hard problem; several works focused on this chal-

lenge could be found in the literature.

We recall that 12 different sessions are recorded

for each person in MOBIO. We can notice in Table

2 that τ, when it is computed for each session sepa-

rately (training and test utterances coming from the

same session), is almost equal to 1, this corresponds

to the ﬁrst 12 cases. On the other hand, τ of the ’all

sessions’ case (majority of utterances come from dif-

ferent sessions) is signiﬁcantly low, see the last line in

Table 2. This observation proves that after achieving

data selection, the majority of the identiﬁcation fail-

ures are due to the multi-session problem. Therefore,

session compensation is still needed to address this

problem.

7 PRIOR WORK

In this paper, we attempted to extend our previous

work (Tikourt et al., 2015) dealing with MFCC se-

lection. In short, the major improvements are sum-

marized in Table 3.

Table 3: Improvements.

previous work current work

features MFCC MFCC, PLP,

∆MFCC

sel. criteria C

,..,C

classiﬁer SVM SVM, KNN

# speaker 10 20

# training utter. 10 10

# test utter. 10 30

session exploration no yes

In addition to SVM, we propose to apply the KNN

classiﬁer, because this last one is non-parametric, ro-

bust and able to separate classes with non-linear com-

plex boundaries. The suggested improvements aim

partly to consolidate results about MFCC selection

obtained in our previous work (Tikourt et al., 2015).

8 CONCLUSIONS AND FUTURE

DIRECTIONS

In this paper, we have described a set of criteria pro-

posed to select the most relevant short-term feature

vectors for the text-independent recognition task.

A universal background model is used for rep-

resenting the speaker-independent information, and

hence it can be used as a framework for the selec-

tion purpose. The general idea consists in selecting

vectors having a low likelihood in the UBM class.

Speaker identiﬁcation experimental tests on the

MOBIO corpus are presented. Results show that the

relevant speaker information is contained in less of

20% of data maximizing the criteria. Not only us-

ing the vectors that minimize the criteria increases

the complexity in time and space, but also reduces the

identiﬁcation rate.

The ﬁndings of this study show that the distance

Euclidean from the UBM center, and the minus-

likelihood in the one-component Gaussian of UBM

are efﬁcient. This supports the idea that UBM has ap-

proximately a hyper-spherical distribution form, such

as a multivariate normal distribution with equal vari-

ances.

It is obvious that an efﬁcient frame pruning speeds

up the recognition computational process since a

small percentage of data is kept for speakers model

matching. Nevertheless, the recognition performance

SIGMAP 2017 - 14th International Conference on Signal Processing and Multimedia Applications

is enhanced only if a good trade-off between the

frame pruning and the speaker modeling is made. On

the one hand, pruning the irrelevant frames causes a

loss in speaker information, but it makes easier the

task of ﬁtting the speakers’ models to data. On the

other hand, using all the frames preserves the entire

speaker information, but it makes the model estima-

tion inaccurate and more complex. This work and

the review of the literature have led us to conclude

that for TISAR systems an efﬁcient frame pruning,

if it is combined with a suitable modeling, may speed

up signiﬁcantly the recognition task without too much

compromising (even improving) the accuracy. In this

optic, frame pruning is an important approach to de-

sign real-time TISAR systems.

The majority of related works attempt to remove

some kinds of the irrelevant frames using speciﬁc cri-

teria based on the silence, the noise, the phonetic

information, or the correlation between successive

frames. The main contribution of this work consists

in applying the UBM model to prune all the irrelevant

frames at once whatever the kind.

To further our research we plan to use this ﬁnd-

ing inside a veriﬁcation TISAR system. Moreover,

developing efﬁcient frame pruning techniques could

be used as a basis to propose new features (e.g super-

vectors) or models by setting more of importance to

vectors maximizing the selection criteria.

REFERENCES

Almaadeed, N., Aggoun, A., and Amira, A. (2016).

Text-independent speaker identiﬁcation using vowel

formants. Journal of Signal Processing Systems,

82(3):345–356.

Beigi, H. (2011). Fundamentals of Speaker Recognition.

Springer Publishing Company, Incorporated.

Benyassine, A., Shlomot, E., Su, H. Y., Massaloux, D.,

Lamblin, C., and Petit, J. P. (1997). Itu-t recommenda-

tion g.729 annex b: A silence compression scheme for

use with g.729 optimized for v.70 digital simultaneous

voice and data applications. Comm. Mag., 35(9):64–

73.

Besacier, L. and Bonastre, J. F. (1998a). Frame pruning for

speaker recognition. In Acoustics, Speech and Signal

Processing, 1998. Proceedings of the 1998 IEEE In-

ternational Conference on, volume 2, pages 765–768

vol.2.

Besacier, L. and Bonastre, J. F. (1998b). Time and fre-

quency pruning for speaker identiﬁcation. In Proceed-

ings. Fourteenth International Conference on Pat-

tern Recognition (Cat. No.98EX170), volume 2, pages

1619–1621 vol.2.

Campbell, W. M., Sturim, D. E., and Reynolds, D. A.

(2006). Support vector machines using GMM super-

vectors for speaker veriﬁcation. IEEE Signal Process.

Lett., 13(5):308–311.

Dehak, N., Kenny, P., Dehak, R., Dumouchel, P., and Ouel-

let, P. (2011). Front-end factor analysis for speaker

veriﬁcation. Audio, Speech, and Language Process-

ing, IEEE Transactions on, 19(4):788–798.

Eatock, J. P. and Mason, J. S. D. (1994). A quantita-

tive assessment of the relative speaker discriminating

properties of phonemes. In Proceedings of ICASSP

’94: IEEE International Conference on Acoustics,

Speech and Signal Processing, Adelaide, South Aus-

tralia, Australia, April 19-22, 1994, pages 133–136.

Khoury, E., Vesnicer, B., and Franco-Pedroso, e. a. (2013).

The 2013 speaker recognition evaluation in mobile en-

vironment. Idiap-RR Idiap-RR-32-2013, Idiap.

Kinnunen, T., Karpov, E., and Franti, P. (2006). Real-time

speaker identiﬁcation and veriﬁcation. IEEE Trans-

actions on Audio, Speech, and Language Processing,

14(1):277–288.

Kinnunen, T. and Li, H. (2010). An overview of text-

independent speaker recognition: From features to su-

pervectors. Speech Commun., 52(1):12–40.

Loizou, P. C. (2013). Speech Enhancement: Theory and

Practice. CRC Press, Inc., Boca Raton, FL, USA, 2nd

edition.

McCool, C., Marcel, S., Hadid, A., and et al. (2012). Bi-

modal person recognition on a mobile phone: us-

ing mobile phone data. Idiap-RR Idiap-RR-13-2012,

Idiap.

McLaughlin, J., Reynolds, D. A., and Gleason, T. P. (1999).

A study of computation speed-ups of the gmm-

ubm speaker recognition system. In EUROSPEECH.

ISCA.

Ramrez, J., Segura, J. C., Bentez, C., Torre, A. D. L., and

Rubio, A. (2004). Efﬁcient voice activity detection al-

gorithms using long-term speech information. Speech

Communication, 42:3–4.

Reynolds, D. A., Quatieri, T. F., and Dunn, R. B. (2000).

Speaker veriﬁcation using adapted gaussian mixture

models. In Digital Signal Processing, page 2000.

Tikourt, A., Rouigueb, A., and Djeddou, M. (2015). Ef-

ﬁcient data selection criteria for speaker recognition.

In 3rd Inter. Conf. on Signal, Image, Vision and their

Applications, Guelma, Algeria.

Frame Selection for Text-independent Speaker Recognition