Local and Global Feature Selection for Prosodic Classification of the

Word’s Uses

Abdenour Hacine-Gharbi

, Philippe Ravier

and François Némo

LMSE Laboratory, University of Bordj Bou Arréridj, Elanasser, 34030 Bordj Bou Arréridj, Algeria

PRISME Laboratory, University of Orleans, 12 rue de Blois, 45067 Orleans, France

Laboratoire Ligérien de Linguistique, University of Orleans, Orleans, France

{philippe.ravier, francois.nemo}@univ-orleans.fr, gharbi07@yahoo.fr

Keywords: Prosodic Classification, Local and Global Prosodic Features, Dimensionality Reduction, Curse of

Dimensionality, Filter Feature Selection, Mutual Information, Classification of Word’s Uses.

Abstract: The aim of this study is to evaluate the ability of local or global prosodic features in achieving a classification

task of word’s uses. The use of French word “oui” in spontaneous discourse can be identified as belonging to

the class “convinced (CV)”or “lack of conviction (NCV)”. Statistics of classical prosodic patterns are

considered for the classification task. Local features are those computed on single phonemes. Global features

are computed on the whole word. The results show that 10 features completely explain the two clusters CV

and NCV carried out by linguistic experts, the features having being selected thanks to the Max-Relevance

Min-Redundancy filter selection strategy. The duration of the phoneme /w/ is found to be highly relevant for

all the investigated classification systems. Local features are predominantly more relevant than global ones.

The system was validated by building classification systems in a speaker dependent mode and in a speaker

independent mode and also by investigating manual phoneme segmentation and automatic phoneme

segmentation. In the most favorable case (speaker dependent mode and manual phoneme segmentation), the

rate reached 87.72%. The classification rate reached 78.57% in the speaker independent mode with automatic

phoneme segmentation which is a system configuration close to an industrial one.

1 INTRODUCTION

Speech communication is a rich but complex way for

exchanging information from a speaker to a listener.

Indeed, among other modalities, speech brings with

itself the information of intonation or prosody. Many

studies highlighted the significant role of prosody in

syntactic, semantic and pragmatic interpretation

(Kompe, 1997) as well as automatic speech

recognition (Wang, 2001), speaker identifica-tion

(Manganaro et al., 2002), language recognition (Mary

and Yegnanarayana, 2008). More recently, since

prosody can help refining semantics of the messages,

this new information modality has found great

interest for semanticists who started to consider that

the prosodic features could play a key role in the

interpretation and classification of different word’s

uses (Petit, 2009). Prosodic classification can find a

particular interest in oral surveys and opinion polls

industry where prosody delivers information that

written language cannot. The goal is to exploit this

finer information given by key-words in an automatic

way for the processing of the large databases at stake.

In a previous paper (Hacine-Gharbi et al., 2015),

preliminary results were obtained as an attempt to

automatically classify the uses of the French “oui” in

a class label “conviction” (CV) or class label “lack of

conviction” (NCV). To that aim, a small vocal

questionnaire inspired from opinion polls has been

created and allowed to collect 118 occurrences for

both classes of ‘oui’. The classification was

performed by a prosodic feature extraction stage

combined with a feature selection stage with a

“wrapper” strategy. The classifier uses hidden

Markov models (HMM) which inputs are the

prosodic feature vectors. Each vector is composed of

the energy and pitch with the dynamic parameters

delta and delta-delta.

The purpose of this work is (i) to enrich the word

representation by the individual word duration; (ii) to

evaluate how the features are relevant for explaining

the two clusters CV and NCV of the full database

manually carried out by the linguistic experts; (iii) to

investigate the role of local and global prosodic

parameters in the classification process. The

motivation for the inclusion of this new parameter

Hacine-Gharbi, A., Ravier, P. and Nemo, F.

Local and Global Feature Selection for Prosodic Classiﬁcation of the Word’s Uses.

DOI: 10.5220/0006251407110717

In Proceedings of the 6th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2017), pages 711-717

ISBN: 978-989-758-222-6

711

comes from the fact that this parameter can be a good

discriminator for the two considered classes. This

parameter cannot be directly included in a HMM

based classifier which requires a sequence of prosodic

feature vectors describing local behaviors. This

means that the HMM based classifier must be

changed which motivates the point (ii). Words have

then to be represented by a single vector instead of a

sequence of vectors making it possible to introduce

the word duration within the single prosodic feature

vector. We therefore consider statistics of the

prosodic feature vectors. These statistics are the

classical mean and standard deviation computed on

the feature vectors sequence for each component.

However, this can be a sensible strategy only if data

are stationary and this is not the case for single words.

So the idea is to consider the so-called local features

which are computed on the quasi-stationary

phonemes which constitute each word. The word

“oui” can be decomposed in the phonemes /w/ and /i/.

The feature vector is thus the concatenation of the

prosodic feature statistics of both phonemes

augmented by their respective duration.

Moreover, we are also interested to know whether if

the local strategy is relevant. We thus also include the

global features, i.e. the prosodic statistics computed

on the whole word, in order to evaluate the effect of

local vs global prosodic features strategy in the

classification system. The answer to this question can

be obtained by performing a feature selection

procedure which will point out either local or global

features within the most relevant ones. Additionally,

the proposed study makes the feature selection

procedure becoming mandatory because a rather

small database size tends to make the rate of

classification drop with the increasing vector size.

The peaking phenomenon suggests that an optimal

number of features have to be selected depending on

the size of the database and on the selected features

as well. To avoid the well known curse of

dimensionality phenomenon (Jain et al., 2000),

dimensionality reduction is necessary. The reduction

can be achieved thanks to the combination of

“greedy” algorithms useful for the feature selection

and mutual information measure which is a powerful

tool for evaluating the shared information between

variables.

The paper is organized as follows: in section 2, the

prosodic classification system of the word “oui” is

proposed and the construction of the statistical

prosodic features vector is detailed. In section 3, the

procedure that selects the relevant features for a

classification task is presented. Section 4 presents

experiments and results in terms of classification rates

and selected features using our database. Section 5

concludes the paper.

2 PROSODIC CLASSIFICATION

OF THE WORD “OUI”

2.1 Classification System

In the present work, the prosodic pattern is a vector of

statistical features (mean, standard deviation)

belonging to the class CV or NCV. An appropriate

powerful classifier for binary classification is a

support vector machine (SVM) classifier. Figure 1

presents the outline of our automatic classification

system of prosodic patterns into word’s use.

A SVM based classification system is classically

composed of two distinct phases, the training phase

and the testing phase. The database is therefore split

into a training database and a testing database. Both

phases rely on a prosodic analyzing step which

consists in transforming the temporal signal of word

“oui” into a sequence of prosodic features vectors.

Then, each sequence is transformed into one vector of

statistical features (mean, standard deviation).

First, during the training phase, the system learns

occurrences of the training database: the set of

statistical features of the training corpus is used for

estimating the support vector parameters of the SVM

classifier. The “svmtrain” MATLAB command is

used with the Kernel function fixed as “rbf”

(Gaussian Radial Basis Function kernel). This

command uses an optimization algorithm to find

support vectorss



, weights w



, and bias B that are

used to classify unknown prosodic vector x.

Second, during the testing phase, the set of statistical

features of the training corpus are the input of the

SVM classifier that gives the recognized class using

the decision equation (command “svmclassify”):

h=w



(



)



where K is the kernel function. If h≥0, then x is

classified as a CV class, otherwise it is classified as

NCV class. The final result is an SVM model that

discriminates the two classes through their statistical

feature vector. The quality of the classification

system is evaluated by a classification rate defined as

TC=





, where N is the total number of occurrences

given at the input of the classifier and S is the number

of misclassified occurrences. The description of the

prosodic features is now given in the next section.

ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods

712

Figure 1: Automatic prosody based classification system

into word’s use.

2.2 Prosodic Feature Extraction

Typical features that characterize prosody are the

energy ∆ (dB) and the pitch F



(Hz). Thanks to

PRAAT software (Boersma and Weenink, 2014),

these parameters are computed every 10 ms on 30 ms

analyzing windows of the temporal speech signal

corresponding to an occurrence of the word “oui”.

A dynamic description of these static parameters



and E is added by computing differential

parameters of first and second order ∆ and ∆∆ using

HTK (Hidden Markov Model Toolkit) library (Young

et al., 1999). Thus, each occurrence of the word “oui”

is represented by a sequence of vectors with 6

prosodic components noted as

E, F



,∆E,∆F



,∆∆E,∆∆F



. Then the statistical

features (mean and standard deviation) of each of the

six features are estimated from the sequence. The

number of features is therefore 13 when also

considering the duration of the sequence. Now, we

define the sequences for three segments: the two local

phonetic segments of phonemes /w/ and /i/ plus the

global segment of the word “oui”. The concatenation

of the segment vectors produces a statistical prosodic

vector composed of 39 features. This vector describes

each word occurrence and this is the input of the

classification system. The output is the class label CV

or NCV which is known during the training phase and

induced during the testing phase.

However, the construction of such 39 prosodic

statistical features vector basically requires the

detection of the words “oui” as well as the

segmentation of each detected word in the two

phonemes /w/ and /i/. These tasks have been manually

achieved with the help of two software: PRAAT for

the detection of words from spontaneous speech

discourses and Easyalign for the phonetic

segmentation. In this work, 115 words have been

segmented by Easyalign and manually corrected.

These steps of detection and segmentation can be

achieved using HMM segmentation but we have

preferred a manual segmentation for guarantying the

best segmentation. Indeed, the best accuracy of some

measures like duration of the word or phonemes is

also needed as new local and global features in this

study. However, in order to be as near as possible

from a real classification system, we investigated an

automatic segmentation system based on the isolated

words we have constructed. The performance of both

classifications systems will be discussed in section

IV. Note also from fig. 1 that the system mentions a

“prosodic feature selection” stage for the training

phase whereas and a “selected prosodic features”

stage is mentioned for the testing phase. This means

that training is useful for selecting relevant prosodic

features used for the testing phase. The feature

selection procedure may either belong to the

“wrapper” strategy or to the “filter” strategy. With the

“wrapper” strategy, the feature selection is performed

upon a classification performance rate criterion. This

necessitates designing a new classifier each time a

feature vector structure is tested. This is a prohibitive

task when the database is growing or when the

number of candidate vectors subsets to be selected is

very large. Since the application of this work

concerns 39 features giving rise to the peaking

phenomenon and necessitating the feature selection),

we therefore prefer the “filter” strategy that permits

to select the features that best describe the classes.

Indeed, the combinatorial possibilities of candidate

subsets are very huge. The “filter” strategy is only

based on the information given by the feature vector

of each occurrence with its belonging class without

having to consider any classification method as

required for “wrapper” strategy. The feature selection

stage is described in the next section.

3 PROSODIC FEATURE

SELECTION

Feature selection consists in choosing a subset S



k features {Y









⋯,Y





}from a set F of n features





⋯,Y



} such that S



keeps most of the

information useful for a classification task. The

quantity of information brought by the subset S



often evaluated thanks to the mutual information (MI)

measure because of its ability of assessing the

nonlinear statistical dependency between variables

(Cover and Thomas, 1991). So the subset S



that

best describes the classes C (in our cases the class

indexes labeled CV or NCV) is the subset that

maximizes the MI between S



andC:



=argmax

⊂

I(C;S). (1)

Local and Global Feature Selection for Prosodic Classiﬁcation of the Word’s Uses

713

To circumvent the prohibitive search which becomes

intractable when the size of S grows, “greedy

forward” search strategies can be employed. The

search is an iterative algorithm that proposes at each

iteration j the best feature Y





from the unselected

features set. This new selected feature is then

appended to the already selected subsetS



generatingS







∪S



(Brown et al., 2012):

Y





=arg max





∈



IC;Y





 (2)

Equation (2) can also be expanded in a multivariate

MI of order 3 between C,Y



and S



as:





=arg max





∈



I

(

C;Y



)

−I



C;Y





 (3)

The evaluation of I



C;Y





 becomes very

difficult when j grows because this evaluation

requires the estimation of high-dimensional

probability density functions that cannot be precise

enough for fixed database sizes. Most of the

algorithms propose a simplification of (3) following

different strategies like MIM, MIFS, MRMR, CMI,

DISR, CIFE, TMI, ICAP (Brown et al., 2012). In

(Brown et al., 2012), the authors conclude that the

JMI strategy provides a good compromise between

precision, flexibility and stability when the database

is small size. They also point out the MRMR and CMI

strategies that better perform than other ones in terms

of balance between high relevance and small

redundancy. We give below the derivation of (3) for

the three selected strategies (Brown et al., 2012).

 MRMR (Max-Relevance Min-Redundancy)





=arg max





∈



I

(

C;Y



)

−

j−1

IY















(4)

 JMI (Joint Mutual Information)





=arg max





∈



I

(

C;Y



)

−

j−1

I



C;Y















(5)

 CMI (Conditional Mutual Information)





=arg max





∈



I

(



)

−max







∈





C;



;







(6)

For JMI and CMI strategies, the term I



C;Y







is

actually computed as IY







−IY







\C. The

MI I(X;Y) between variables X and Y is expressed as

(

X;Y

)

∬



(

,

)





(

,

)



(



)



(



)







where



(

,

)

is the joint distribution of

(

,

)

and



(



)

and

(



)

are the marginal distributions . This

continuous definition can be estimated by

discretization of the I(X;Y) formula applying the

histogram low mean square error estimation as

described in (Hacine-Gharbi et al., 2013).

The three strategies have been used and compared for

the feature selection procedure.

4 EXPERIMENTS AND RESULTS

Three experiments are conducted in order (i) to

evaluate the relevance of the type of features, either

local or global; (ii) to know whether a set of selected

features is able to completely explain the two clusters

composed of 64 CV words and 51 NCV words

labeled by the linguistic experts; (iii) to evaluate the

classification rates in real situations where feature

selection and training phase are performed on the split

database. In this last case, the phonetic segmentation

has been achieved either manually by visual

inspection or automatically by modeling the word

‘oui’ into the phonemes /w/ and /i/. Each phoneme is

modeled by a 3 states HMM. Each state is modeled

by a Gaussian mixture model (GMM) with three

Gaussians. The implementation of the system is

carried out using the HTK library (Young et al.,

1999). This task helps identifying the begin and the

end of each phoneme which permits to calculate the

local parameters in the phoneme regions. The

performance of this segmentation is compared to the

performance obtained with a manual segmentation in

terms of precision or classification rate.

4.1 Performance Study using Local and

Global Features

This experiment permits to analyze the pertinence of

local or global prosodic features that should be able

to explain the index variable representing the class

each occurrence belongs to. The same database of 115

occurrences was considered both for the training

phase and the testing phase. The goal in this section

is to evaluate whether the features are pertinent for

explaining the two clusters of CV and NCV words

proposed by the linguistic experts. The performance

study is achieved by independently considering three

sets of features: local features are referred to the

statistical prosodic features for the phoneme /w/

ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods

714

Table 1: Performance comparison of global (GF) and local features (LFi and LFw).













∆



∆





∆∆



∆∆















∆



∆





∆∆



∆∆





LFw 70.44 78.26 79.13 87.82 93.04 92.17 93.91 93.91 96.52 97.39 98.26 98.26 98.26

LFi 60.87 71.30 74.78 81.74 85.22 84.35 86.09 87.83 91.30 87.83 87.83 86.09 89.57

GF 58.26 73.91 73.91 80.87 86.08 88.70 93.04 94.78 95.65 95.65 96.52 95.65 96.52

Table 2: Classification rates obtained with feature selection procedure and the selected feature names at each iteration.



            

MRMR 76.52 80.87 88.70 88.70 93.04 94.78 95.65 98.26 99.13 100 100 100 100

()

∆



()



()

∆∆

()





()

∆∆



()

∆∆

()







()

∆∆

()



()

∆∆

()

∆



()

JMI 76.52 80.87 82.61 84.35 80.87 93.91 93.04 93.91 93.91 96.52 96.52 96.52 96.52

()

∆



()



()

∆∆



()



()

∆∆

()

∆

()

∆

()

∆∆



()

∆



()





()





()

CMI 76.52

()

80.87

∆



()

88.70



()

86.96



87.83

∆∆

()

88.70

∆

()

93.04





()

93.04

∆∆



()

96.52

∆



()

96.52

∆∆



()

97.39

∆



()

98.26



()

98.26

∆

()

(LFw) or the phoneme /i/ (LFi) and global features

are referred to the whole word “oui”. The results

expressed in terms of the classification rate are shown

in Table I, in the natural order of the 13 parameters

composed of 6 mean values



,μ





,μ

∆

,μ

∆



,μ

∆∆

,μ

∆∆



and 6 standard

deviation values σ



,σ





,σ

∆

,σ

∆



,σ

∆∆

,σ

∆∆



plus

the duration d of the segment. This table highlights

the pertinence of the local features corresponding to

the first phoneme /w/ since values are always higher

for LFw. The LFi values are less informative for the

classification task since the classification rates are in

average 8% below those obtained for LFw. However,

the 13 available features do not explain 100% of the

class indexes. It is thus necessary to increase the

number of features by concatenating them.

4.2 Feature Selection Results

This part is devoted to the selection of the most

relevant features explaining the classes or clusters CV

and NCV among the 39 local and global features. The

complete database with the 115 occurrences is

employed for the selection. The classification rates

are evaluated on the same database. Results are

reported in table II for the 13 first selected features.

Each reported classification rate mentions the new

selected feature by one of the three tested strategies

(MRMR, JMI or CMI). The features are mean,

standard deviation  or duration. Superscript

conventions are adopted for indicating a local or

global feature: the notations

() and () are referred to

a local feature for phonemes /w/ and /i/ respectively;

the notation

 is referred to a global feature.

The results in table II show that the 10 first features

selected from 39 by the MRMR strategy can explain

both clusters manually carried out by the linguistic

experts since the classification rate is 100%.

Moreover, the MRMR strategy always gives the best

results comparatively to JMI and CMI. The most

interesting result is the selection of the duration of the

phoneme /w/ as first feature, whatever the strategy. In

addition, the dynamic pitch feature of the second

phoneme is a major feature since always selected as

second one. The pitch is relevant when observed

through its temporal domain variations expressed by



∆



()

. The results also show that the global duration

parameter is always selected among the 10 first

parameters whatever the strategy. Note an other

important result suggesting that local features are

more relevant than global ones since the 10 first

selected features are predominantly local.

We analyzed in table III the role of the segment

duration features (DF). Classification rates are shown

with single feature and the combinations of two

features from 3, plus the 3 features. The role of



()

confirmed as first feature. Moreover, the combination

of two duration features also requires



()

to be

selected. The conclusion is that this parameter is

highly relevant.

Table 3: Classification rates obtained with duration

features.

{

()

} {

()

}



}

{

()

()

}

{

()



}

{

()



}

{



()

,

()

,



}

DF 76.52 64.35 70.43 85.22 80.86 85.22

Local and Global Feature Selection for Prosodic Classiﬁcation of the Word’s Uses

715

4.3 Classification Rates in a Real

Context

The classification has been evaluated on dependent

speaker mode and on independent speaker mode. In

the first mode, the experiment evaluates the

classification rates when classically splitting the

database in a training part (58 occurrences, 27 CV and

31 NCV) and a testing part (57 occurrences, 37 CV

and 20 NCV). The occurrences of each speaker were

dispatched into both parts (speaker dependent

system). Two classification systems were built: the

first one investigated a manual phoneme segmenta-

tion achieved by linguistic experts while the second

one investigated automatic phoneme segmentation.

For each system, the classification rates are displayed

in Fig. 2 and Fig. 3 as a function of the number of

selected features, using one strategy among MRMR,

JMI or CMI.

Figure 2: Classification rates with feature selection

achieved on the 58 occurrences with manual segmentation

in speaker dependent mode.

Figure 3: Classification rates with feature selection

achieved on the 58 occurrences with automatic

segmentation in speaker dependent mode.

Both graphs show that the peaking phenomenon

which is due to the relatively too small number of

occurrences for the training phase (58) regarding the

potential total number of features (39). The best

results give a classification rate of 87.72% with 9

features and MRMR strategy in the manual segmenta-

tion case and 84.21% in the automatic segmentation

case with 7 features and CMI strategy, respectively.

The three selection methods gave the same first

selected feature (duration of the /w/ phoneme),

whatever the segmentation procedure. In both cases,

the feature selection process permits to limit the

number of features to be considered for classification.

In the second mode, the experiment evaluates the

classification rates when classically splitting the

database in a training part (59 occurrences, 33 CV and

26 NCV) and a testing part (56 occurrences, 31 CV

and 25 NCV).

In this mode, the speakers participating in the testing

phase have not participated in the training phase. Two

classification systems were similarly built to the

independent case. For each system, the classification

rates are displayed in Fig. 4 and Fig. 5 as a function

of the number of selected features, using one strategy

among MRMR, JMI or CMI.

Figure 4: Classification rates with feature selection

achieved on the 59 occurrences with manual segmentation

in speaker independent mode.

Figure 5: Classification rates with feature selection

achieved on the 59 occurrences with automatic

segmentation in speaker independent mode.

0 5 10 15 20 25 30 35 4

Feature number

Classification rate

MRMR

JMI

CMI

0 5 10 15 20 25 30 35 40

Feature number

Classification rate

MRMR

JMI

CMI

0 5 10 15 20 25 30 35 40

Feature number

Classification rate

MRMR

JMI

CMI

0 5 10 15 20 25 30 35 40

Feature number

Classification rate

MRMR

JMI

CMI

ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods

716

Like in the dependent mode, the three selection

methods highlight the /w/ phoneme duration as a

relevant feature since all the methods firstly select

this feature. The maximum rate for the manual

segmentation (85.71% with JMI strategy and 8

features) is near from the rate previously achieved in

the dependant mode. As expected, a classification rate

decrease is observed in the independent mode with

automatic segmentation which however remains

acceptable when considering this system close from

an industrial system (maximum of 78.57% whatever

the feature selection method).

Nevertheless, these results remain partial ones since a

real functional system requires a large database which

was not available in our study.

5 CONCLUSION

The aim of this study was to evaluate the ability of

local or global prosodic features for explaining the

two clusters carried out by linguistic experts and for

classifying the word’s uses of “oui” in a real context

of spontaneous discourse. The word’s uses were

identified as belonging to the “convinced” or “lack of

conviction” class. The results showed that 10 features

are sufficient to fully explain both clusters CV and

NCV based on the 115 occurrences of a self made

corpus which have been labeled by linguistic experts;

the features having being selected thanks to the

MRMR filter selection strategy. The 10 relevant

selected features by this strategy are local for the most

part. All the results showed that the first relevant

feature was the /w/ phoneme duration. The system

was validated by building classification systems in a

speaker dependent mode and in a speaker

independent mode and by also investigating manual

phoneme segmentation and automatic phoneme

segmentation. In the case of speaker dependent mode

and manual phoneme segmentation, the rate reached

87.72%. The classification rate reached 78.57% in the

speaker independent mode with automatic phoneme

segmentation which is a system configuration close to

an industrial one. These results are partial and

preliminary ones regarding the size of the database.

However, these are promising for industrial

applications like automatic processing of large

database oral opinion polls.

ACKNOWLEDGMENT

This work is funded by the région Centre Val de

Loire, France. This collaborative work implies four

laboratories of the University of Orléans (LLL,

PRISME/IRAuS, LIFO, MAPMO). All the persons

involved in this project are acknowledged for their

active participation.

REFERENCES

Boersma, P. and Weenink, D., 2014. Praat: doing phonetics

by computer. [Online] (5.3.75) Available at:

www.praat.org [Accessed 2014].

Brown, G., Pocock, A., Zhao, M.-J. and Lujan, M., 2012.

Conditional Likelihood Maximisation: A Unifying

Framework for information theoretic feature selection.

Journal of Machine Learning Research, 13(1), pp.27-

66.

Cover, T. and Thomas, J., 1991. Elements of information

theory. New York: Wiley Series in telecommunica-

tions.

Hacine-Gharbi, A. et al., 2013. A new histogram-based

estimation technique of entropy and mutual informa-

tion using mean squared error minimization. Computers

and Electrical Engineering, 39(3), pp.918-33.

Hacine-Gharbi, A., Petit, M., Ravier, P. and Nemo, F.,

2015. Prosody Based Automatic Classification of the

Uses of French ‘oui’ as Convinced or Unconvinced

Uses. In 4th International Conference on Pattern

Recognition Applications and Methods (ICPRAM).

Lisboa, Portugal, 2015.

Jain, A., Duin, R. and Mao, J., 2000. Statistical pattern

recognition: a review. Trans. Pattern Analysis and

Machine Intelligence, 22, (1), pp.4-37.

Kompe, R., 1997. Prosody in Speech Understanding

Systems. LNAI.

Manganaro, L., Peskin, B. and Shriberg, E., 2002. Using

prosodic and lexical information for speaker. In

ICASSP., 2002.

Mary, L. and Yegnanarayana, B., 2008. Extraction and

representation of prosodic features for language and

and speaker recognition. Speech Communication,

pp.782–96.

Petit, M., 2009. Discrimination prosodique et

représentation du lexique: application aux emplois des

connecteurs discursifs. thesis. PhD Thesis, University

of Orléans.

Wang, C., 2001. Prosodic modelling for improved speech

recognition and understanding. PhD Thesis,

Massachussetts Institute of Technology.

Young, S., Kershaw, D., Odell, J. and Ollason, D., 1999.

The HTK Book. Cambridge: Entropic Ltd.

Local and Global Feature Selection for Prosodic Classiﬁcation of the Word’s Uses

717