A Better Understanding of Esophageal Speech Excitation Source

Behavior

Radhouane Bouazizi and Soﬁa Ben Jebara

Lab. COSIM, Ecole Sup

erieure des Communications de Tunis, Carthage University,

Route de Raoued 3.5 Km, Cit

e El Ghazala, Ariana 2083, Tunisia

Keywords:

Esophageal Speech, Excitation Source, Esophagus Extremity Vibration, Opening/Closing Cycle.

Abstract:

Understanding the excitation source of the esophageal speech is a key approach for understanding the

esophageal speech. In this paper, we extract the excitation source using an inverse ﬁltering approach and

we analyze it. We, for example, show some similarities with an artiﬁcial EGG signal. We also detect the clos-

ing instants in order to deﬁne cycles of opening/closing of esophagus extremity and to recognize the equivalent

of glottal cycles. These cycles are classiﬁed into different types according to their characteristics. A physical

explanation of the esophagus extremity behavior is systematically given at the different steps of the analysis.

1 INTRODUCTION

For several reasons that range from innate, pathologi-

cal (cancer for example) to accidental reasons, a per-

son, namely called laryngectomee, may lose his voice

after a laryngectomy operation (vocal cords eradica-

tion) (Brown et al., 2003). Therefore, the person be-

comes unable to produce speech in a normal manner.

This is explained by the need of the vocal cords in

the process of speech production. Esophageal voice

consists in producing speech by bringing air and re-

leasing it through the end of the esophagus which re-

places the vocal cords when speaking in normal man-

ner. One kind of esophageal speech production mech-

anism begins by injecting the air through the mouth in

order to reach the extremity of the esophagus. While

leaving, this trapped air causes the vibration of the

tissue at the entrance of the esophagus. Then, as in

the case of normal speech, where the role of the vocal

cords is to vibrate under the action of the pressurized

air ejected from the lungs, the esophagus extremity

does the same thing with the incoming pressured air.

The signal created at this level of speech production

mechanism is called the excitation source and have a

crucial role in producing substitution voice of good

quality. In fact, all the remaining of the mechanism is

the same as the one of normal speech (modulation in

the vocal tract and radiation through the lips).

This paper aims to study in depth the excitation

source of esophageal voice and the behavior of the

end of the esophagus. In fact, contrary to vocal cords

whose behavior is well studied and understood, the

esophagus extremity behavior, as excitation source, is

still not well mastered.

The paper is organized as follows. In section 2,

we will show the complexity of signals (esophageal

speech and its excitation source) in the temporal and

frequency domains. Section 3 is devoted to establish

an equivalence between an artiﬁcial ElectroGlottoG-

raphy signal and the source of the esophageal speech.

In section 4, we deﬁne and localize the closure in-

stants of the esophagus extremity. In section 5, differ-

ent types of opening/closing cycles are identiﬁed and

characterized. Finally, a conclusion is given.

2 SHOWING ESOPHAGEAL

VOICE PARTICULARITY

In this section, we will ﬁrst, describe the speech

model that is appropriate for esophageal speech.

Then, a comparison between natural and esophageal

voices and their excitation sources, in time and fre-

quency domains, is made.

2.1 Source/Filter Model of the Speech

Fig. 1 illustrates an example of the signals and the ﬁl-

ters appearing in the source/ﬁlter model of esophageal

voice. The excitation source g(n) is shown in the

temporal and frequency domains. It is then ﬁltered

163

Bouazizi R. and Ben Jebara S..

A Better Understanding of Esophageal Speech Excitation Source Behavior.

DOI: 10.5220/0004228301630168

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2013), pages 163-168

ISBN: 978-989-8565-36-5

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: Illustration of source/ﬁlter model of esophageal speech.

by the oral and nasal cavities ﬁlter V ( f ) whose fre-

quency response is shown. Finally, the voice is pro-

duced via lips radiation whose transfer function is a

simple derivative operation (L(z) = 1 − z

−1

It is important to note that, given the linearity of

the model and for simplicity reasons, one may con-

sider that the role of the vocal tract and the lips can

be inverted so that the derivative is applied immedi-

ately to the excitation source. This kind of model is

illustrated in the bottom part of Fig. 1.

2.2 Time and Frequency Domains

Representations of Speech

In Fig. 2.a (resp. 2.c) , an example of time domain

representation of natural (resp. esophageal) speech of

the phoneme /a/ is presented. As the periodicity and

regularity are visually observed for natural voice, it is

not the case for esophageal voice where many ﬂuctu-

ations that are difﬁcult to characterize with the naked

eye are observed. Globally, the signal is unstructured

and the periodicity of the signal is not clear.

Fig. 2.b and 2.d show the frequency domain rep-

resentation of the same signals. The natural voice

is characterized by power peaks at frequencies mul-

tiples of the fundamental frequency F

whereas the

esophageal voice does not have such structure. It is

characterized by very marked peaks in the frequency

region around 1 KHz.

2.3 Time and Frequency Domains

Representations of the Excitation

Source

To go further in the comparison between natural and

esophageal speech, we propose to study the excitation

source. It is estimated using the iterative adaptive in-

verse ﬁltering IAIF method described in (Alku et al.,

2004). It works basically as follows: the vocal tract

transfer function model is iteratively estimated. Then

its contribution is cancelled from the speech signal

via the so-called Inverse Filtering to obtain the glottal

ﬂow of a speech signal.

In Fig. 3.a and Fig. 3.c, the glottal waves of the

signals of Fig. 2 are given. The excitation source

of the natural voice has a periodic shape and a well-

ordered structure against a non-periodic and a miss-

ordered structure of the esophageal voice excitation

source.

Fig. 3.b and 3.d shows the frequency domain

representations of the same signals. The one of

esophageal speech is characterized by an energy con-

centration at low frequencies (from 0 to 600 Hz) but

it is not structured in peaks placed at frequency multi-

ples of F

as it is the case of the glottal ﬂow of natural

speech.

As a conclusion, since there is no clear similarity

between the speech signal of the natural voice and the

one of esophageal voice, the remainder of the paper

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

164

(a) (b)

Figure 2: Time (a) and frequency (b) domains representations of the natural speech and time (c) and frequency (d) domains

representations of the esophageal speech.

(a) (b)

Figure 3: Time (a) and frequency (b) domains representations of the excitation source of natural speech and time (c) and

frequency (d) domains representations of the excitation source of esophageal speech.

will be devoted to a deeper analysis of the excitation

source. The purpose is to better understand the behav-

ior of the esophagus extremity, the organ responsible

of esophageal speech production.

3 MATCHING THE EXCITATION

OF THE ESOPHAGEAL

SPEECH WITH AN

ARTIFICIAL EGG SIGNAL

The ElectroGlottoGraphy (EGG) is a non invasive ex-

perimental procedure used to measure the glottal ﬂow

ABetterUnderstandingofEsophagealSpeechExcitationSourceBehavior

165

of natural speech. The acquired signal describes the

vocal folds vibration where period cycles of open-

ing/closing are observed. The top portion of the signal

reﬂects opening of the vocal folds, while the bottom

portion reﬂects closing. It can be inverted according

to the electrodes connection. The EGG signal ﬁts well

the excitation source (see for example Fig.4 for the

example of the natural phoneme /a/).

The question that arises is: is it possible to de-

scribe the esophagus extremity in the same man-

ner? In other words, does the excitation source of

esophageal speech looks like an EGG signal?

But, of course, in case of esophageal speech, there

is no known invasive external device that helps acquir-

ing a signal describing the behavior of the esophagus

extremity in terms of opening/closing cycles.

To overcome this limitation, we propose an alter-

native solution, to create artiﬁcially, the equivalent of

an EGG signal. This procedure is described as fol-

lows.

• Given that the main difference between these two

types of voice resides in the excitation source, we

propose to try to match an EGG signal generated

in an artiﬁcial manner to the excitation source of

the esophageal voice.

• The glottal ﬂow of natural speech follows the well

known LF model (Fant et al., 1985). So, by ﬁxing

the fundamental frequency F

, one can generate

easily an artiﬁcial EGG signal.

• The esophageal speech is not systematically pe-

riodic in all speech segments. So we try to

ﬁnd manually a fundamental frequency value for

which the esophageal speech excitation signal ﬁts

the artiﬁcial EGG.

• The matching is done by locating areas where we

can ﬁt the maxima (resp. minima) of the EGG to

the maxima (resp. minima) of the source.

Fig.5 shows the matching between the artiﬁcial

EGG of a natural voice and the excitation source of

the phoneme /a/ of an esophageal voice. We can see

that we can locate some areas where there is an impor-

tant similarity between the two signals (two examples

of areas are delimited by bold discontinuous vertical

lines). We can also match the position of the max-

ima (resp. minima) of the source to the maxima (resp.

minima) of the EGG. Furthermore, we can locate, be-

tween two successive minima, a kind of period, which

corresponds to a cycle of phonation duration (funda-

mental period). In the same ﬁgure, the two pointed

periods have close values (4.3 ms and 4.15 ms). This

local period allows ﬁnding a frequency that enables

us to generate an artiﬁcial EGG that can ﬁt perfectly

(in a local manner) the excitation source.

Figure 4: A matching between the EGG signal and the ex-

citation source of a natural voice.

Figure 5: A matching between an artiﬁcial EGG signal

(bold dashed line) and the excitation source (dashed line)

of an esophageal voice.

Elsewhere, there are more difﬁculties to perform

such identiﬁcation of previous similarities. In fact,

the ﬂuctuations of the excitation source are so impor-

tant that we can not identify the maxima of the exci-

tation source (see for example circled areas in dashed

ellipses in Fig. 5). We may call these cycles ”deterio-

rated cycles”. For example, from the instant 0.325 ms

to the instant 0.33 ms, we can notice the absence of a

clear maximum.

4 EXTRACTION OF RELEVANT

INSTANTS IN THE

EXCITATION SIGNAL

After using an artiﬁcial EGG signal to identify the

cycles describing the behavior of the esophagus ex-

tremity, let’s look for the signiﬁcant instant of closing

to better identify and characterize cycles within the

esophageal speech signal.

In natural speech, the physical phenomenon of

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

166

Figure 6: A cycle of glottal ﬂow and its derivative.

voiced sounds pronunciation takes its origin in the

lungs. Air passes later through the glottis and ex-

erts a force on the vocal cords making them tick. The

moment when the cords begin to separate is called

Glottal Opening Instant (GOI). Similarly, after a max-

imum separation, the vocal cords tend to return to

their original positions and the moment when the vo-

cal cords re-stick is called Glottal Closure Instant

(GCI). It corresponds to a minimum in the glottal ﬂow

derivative (see Fig. 6).

Is the same thing happens in esophagus extremity

when trying to produce esophageal speech? This is

the problem addressed now.

In order to identify the particular instant of clos-

ing GCI, we exploit the Dynamic Programming Pro-

jected Phase-Slope Algorithm (DYPSA) (Kounoudes

et al., 2007). It is an automatic algorithm which

operates using the speech signal alone without the

need for an EGG signal. It incorporates a technique

based on phase-slope function for estimating GCI

candidates which are the instants of positive-going

zero-crossings in the phase-slope function. Next,

DYPSA algorithm employs dynamic programming to

select the most likely candidates according to a de-

ﬁned cost function. This latter is a combination of

ﬁve sub-functions considering pitch deviation, norm

amplitude Consistency, ideal phase slope function,

speech waveform similarity and projected candidate

(Kounoudes et al., 2007).

When DYPSA algorithm is applied on natural

speech, the GCI instants correspond perfectly to the

instants when the excitation source derivative reaches

the minima. In case of esophageal speech, we notice

that the majority of the detected instants with DYPSA

algorithm correspond to the minima of the excitation

source derivative extracted with IAIF algorithm. But

other candidates are badly placed in other positions

and some candidates are missed (see Fig. 7 for exam-

ple). This constatation can be explained by the fact

there is not systematically complete opening and clos-

ing of the esophagus extremity. Consequently, there

Figure 7: Superposition of GCI instants and the excitation

source derivative.

Table 1: Pseudo-fundamental frequency for different

phonemes of esophageal speech.

Phoneme 1 /a/ Phoneme 2 /a/ Phoneme 3 /o/

106 Hz 185 Hz 208 Hz

113 Hz 220 Hz 172 Hz

106 Hz 178 Hz 185 Hz

105 Hz 204 Hz 204 Hz

108 Hz 190 Hz 169 Hz

is no real cycles and real periods along the whole sig-

nal.

We’ll give more details in next section.

5 IDENTIFICATION AND

CHARACTERIZATION OF

PSEUDO-CYCLES

In the central part of the signal of Fig. 7, some glottal

cycles, deﬁned as the period between two successive

GCI, are detected. Their duration vary from 8.4 ms

to 10.5 ms. They are called ”pseudo-periods” as they

change, even in a smooth way, from one cycle to an

other.

In Tab. 1, we consider different phonemes, namely

/a/ and /o/ and we attribute to some consecutive cy-

cles a value of the fundamental frequency (inverse of

the duration). This table shows that the fundamental

frequency varies systematically from one cycle to its

neighbor. It conﬁrms the fact that the esophagus ex-

tremity vibration is not regular and not perfectly peri-

odic, as it is the case of vocal folds.

In Fig. 8, we were able to match the identiﬁed cy-

cles between two successive GCI to the cycles of an

EGG generated at a frequency equivalent to the aver-

age of the pseudo-periods. The signal considered is

the central part of the one in Fig. 7, whose beginning

and ending are delimited by the the ﬁrst and the last

vertical dashed lines. The average frequency of 108

Hz is considered in Fig. 8 which corresponds to a

pseudo-period of 9.2 ms. We recall that, according to

ABetterUnderstandingofEsophagealSpeechExcitationSourceBehavior

167

Figure 8: Classiﬁcation of cycles according to esophagus

extremity opening/closing.

Fig. 7, the real pseudo-periods range from 8.4 ms to

10.5 ms.

Moreover, it is important to note that, by considering

this average fundamental frequency in Fig. 8, it is

obvious that CGI instants does not correspond to the

minimum of the excitation source derivative, since the

fundamental real period is replaced by the averaged

one.

After a deep study of an important number of

esophageal speech sequences in terms of cycles struc-

ture and duration, one can go further in the analysis

of cycles by deﬁning a kind of cycle’s classiﬁcation.

The proposed classiﬁcation is illustrated in the signal

of Fig.8.

• The two ﬁrst cycles are considered as regular cy-

cles or acceptable as they look more or less similar

to the derivative of the EGG. Moreover, the local

maximum of the cycle is identiﬁed as the maxi-

mum of the derivative of the EGG.

• The third cycle is classiﬁed as non-regular or fa-

tigue cycle. This is a deteriorated cycle during

which we cannot identify its maximum to the

maximum of derivative of the EGG. It means that,

the esophagus extremity was not able to open

completely in easy manner and was trying to do

it more longer in time than during regular cycles.

• The two following cycles behave more similar to

the derivative of the EGG. It means that there is a

return to regular cycles after a fatigue cycle.

In brief, we may explain getting a diversity of be-

havior by the fact that the esophageal voice, compared

to the natural voice, is more difﬁcult to produce and

that the speaker is not ﬂuent in producing this kind of

voice. This may also explain the occurring of the so-

called periods of fatigue during which the esophagus

does not perform a correct cycle of opening/closing.

This time interval of fatigue occurs after few proper

cycles of opening/closing. After this time interval of

fatigue which can be considered as a time interval of

relaxing, the esophagus extremity recovers and pro-

duces more correct cycles. So, it has a behavior that

could be a little recognizable or similar to the behav-

ior of the vocal cords.

6 CONCLUSIONS

Even the complexity of the esophageal speech, the ex-

traction of the excitation source permits to develop an

analysis and a better comprehension of the esophagus

extremity. We have, for example, shown some sim-

ilarities between the EGG and the excitation source.

Moreover, thanks to the localization of the CGI in-

stants, we identiﬁed the equivalent of the glottal cy-

cles in the esophageal speech. The classiﬁcation of

those cycles allowed to describe the status of the

esophagus while producing the speech.

REFERENCES

Alku, P., Story, B., and Airas, M. (2004). Evaluation of

an inverse ﬁltering technique using physical modeling

of voice production. In Proc. of INTERSPEECH. pp.

497-500, Kyoto-Japan.

Brown, D. H., Hilgers, F. J. M., Irish, J. C., and Balm,

A. J. M. (2003). Postlaryngectomy voice rehabilita-

tion: state of the art. In World Journal of Surgery. vol.

27, no.2, pp. 824-831.

Fant, G., Liljencrants, J., and Lin, Q. (1985). A four-

parameter model of glottal ﬂow. In journal STL-

QPSR. vol. 4, pp. 1-13.

Kounoudes, A., Naylor, P. A., Gudnason, J., and Brookes,

M. (2007). Estimation of glottal closure instants in

voiced speech using the dypsa algorithm. In IEEE

Trans. Speech and Audio Processing. pp. 34-43.

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

168