IV/ Evaluation objective de la discrimination
phonétique avec l'implant cochléaire Digisonic®
Pour évaluer de manière objective
l'intelligibilité du signal émis à la partie interne de
l'implant cochléaire Digisonic, trois types de reconnaissance ont
été effectués. Dans un premier temps nous avons
évalué la reconnaissance obtenue par ordinateur. Celle-ci n'a
qu'une valeur indicative sur l'intelligibilité maximale que l'on peut
avoir, puisque le mode de reconnaissance des phonèmes est très
éloigné de celui pratiqué par les sujets implantés
cochléaires. En second lieu, nous avons mesuré
l'intelligibilité du signal de l'implant cochléaire obtenu chez
des normo-entendants et des surdités moyennes en restituant par un
algorithme, le signal de manière acoustique. Pour finir, nous avons
mesuré l'intelligibilité obtenue chez des sujets implantés
cochléaires.
al Reconnaissance des voyelles par analyse
discriminante
Avant de mesurer les performances du sujet implanté
cochléaire, il est intéressant d'estimer l'intelligibilité
du signal émis par l'implant cochléaire. Pour ce faire nous avons
évalué par analyse discriminante (Rouanet et le Roux, 1993), la
séparation et le pourcentage de discrimination d'un échantillon
de 4 voyelles, puis des 16 voyelles de la langue française (cf figure
8).
Une première étude (C. Berger-Vachon, et al, 1997)
a eu quatre objectifs.
1- développer une technique de reconnaissance par une
analyse discriminante (test plus performant que la mesure de distance
euclidienne),
2- évaluer la séparation et la discrimination des
voyelles /a/, /i/, /u/, /3/ via l'implant cochléaire
Digisonic®,
3- comparer les données obtenues dans les 2 modes de
l'implant cochléaire le mode 'A' - i.e. mode parole - (avec au maximum 6
canaux ouverts par cycle) et en mode 'N' - i.e. mode musique - (avec au maximum
15 canaux ouverts par cycle),
4- comparer les performances effectuées par le traitement
de l'implant cochléaire à celles obtenues par un modèle
d'implant simulé sur PC.
Les résultats montrent une grande efficacité de
l'analyse discriminante dans la classification des voyelles /a/, /i/, /u/, /3/.
Le mode 'A' semble être le plus efficace mais il n'est pas
significativement différent du mode 'N' et du simulateur d'implant. Les
conclusions sont peu affirmatives car le nombre de locuteur est faible (2) et
les pourcentages de reconnaissance très proche de 100.
Article 4 :
VOWEL PROCESSING THROUGH A COCHLEAR IMPLANT :
A
model of speech coding
J.C. Bera, S. Gallégo, L Collet, C.
Berger-Vachon
Advances in Modeling & Analysis, 1999, in press
L'objectif de cette étude a été
d'évaluer les séparations d'une population plus importante de
voyelles que l'étude précédente (les 16 voyelles de la
langue française) via l'implant cochléaire
Digisonic® et un modèle simulant l'implant
cochléaire.
L'analyse des résultats montrent de très grands
pouvoirs discriminants des différentes populations de voyelles (0.94
à 1.00). La classification par un modèle basé sur la FFT
est identique voire moins bonne que celle de l'implant
Digisonic®.
Le traitement effectué par l'implant apporte donc
suffisamment d'informations pour permettre de différencier toutes les
voyelles de la langue française par une analyse discriminante.
Vowel Processing Through a Cochlear Implant
A Model
of Speech Coding
J.C. Bera *, S. Gallego **, L. Collet***, C.
Bergpr-Vachon***
* Acoustic Centre, Ecole Centrale de Lyon, 69131 Ecully-CEDEX
(France)
**MXM Laboratories, 06224 Vallauris-CEDEX (France)
***Laboratory
« Perception and Auditory Mechanisms »,
ORL Dpt, E. Herriot
Hospital, 69437 Lyon CEDEX 03 (France)
Abstract
The construction of an efficient code for deaf people fitted with
a Cochlear Implant (CI) is still an op cri problem.
Classically, a spectrum analysis of the acoustical signal is
ma :'e and periodieally distributed at the end- of the auditory nerve of a
patient. Clinically, this approach has been widely used by the physicians Also,
analytical considerations need to be made on the acoustic signal.
In this paper, the results obtained with the discrimination of
the French vowels taken at the output of a CI are discussed and compared with a
FFT analysis (CI &FFT are two models). Vowels are well separated, two by
two, by a discriminant analysis... even those which were expected to be close,
using both strategies. The discussion indicates that the eigenvalues or the
distances used to assess the separation may not be a very relevant item to
simulate the human behaviour in this situation because they separate vowels
which are similar to the ear.
Further studies need to be carried out in order to understand
better the phenomenon.
Key-words: Cochlear implants, Signal
processing, Mathematical strategies, Confusion matrices
Paper presented, and selected, at CCM'98 (Contribution of
Cognition to Modelling), International AMSE-Conference, Lyon-France, 6-8 July
1998.
1.Introduction
The recognition of speech by human beings has raised a lot of
questions [1,2]. Many models have been established and most of the strategies
start with a frequency-time representation of the signal, which is further
processed [3,4,5].
This approach is widely accepted for the ear. In the inner
ear, it is admitted that the cochlea performs a tonotopic decomposition of the
acoustic signal which is distributed at the ends of the auditory nerve. Then,
complex mechanisms occur, starting at the brainstem, to end up into an
interpretation finally given by the brain.
In the case of totally deaf people, the cochlea is
non-functional and it breaks the auditory chain. To beat this handicap, a
cochlear implant is surgically introduced and electrodes are put in the
cochlear duct. Then, electrical pulses are delivered according to a signal
analysis performed by an external device called the speech processor [6].
Mos6 of the studies conducted with patients point out that in
difficult situations, the results stay below 70% [7,8,9]. In the experiments
the stimulus is embedded in standard contexts such as «hvd» for
vowels (where v stands for the vowel) and «aCa» for consonants (where
C indicates the consonant). Most of these studies have deepiy analysed die
situation with the patient, but suffer of a lack of consideration from the
signal processing point of view.
For instance, it is well known that the patient's background
has a tremendous influence on his ability to perform die recognition and use
efficiently his prosthesis. The evolution with time of die recognition scores
has to be taken into account. At die end, it turns out that a deeper look into
die intrinsic properties of the signal should bring interesting considerations
on what is to be expected from die patient. For instance, it is difficult to
imagine that if two stimuli are similar, the patient would separate them in a
nonsense context. But, the similarity is something which is not obvious to be
defined «mathematically speaking». This is why we will take two
approaches to evaluate the distance between «elementary» phonemes.
An objective assessment of the signal processing can be done
directly, and compared to the classical FFT analysis.
In the language, vowels are supposed to be stationary and
their structure can be represented by a stable vector giving, with its
co-ordinates, the distribution of energy in die spectrum. This is a very easy
way to see the signal and interesting to make objective studies on the
signal.
The questions which are raised in this work are:
- what is the efficiency of the coding of the French vowels
by the French cochlear implant (Digisonic DX-10 of MXM) relatively to the
classical coding, obtained in the same conditions, with a FFT analysis?
-what is the performance of both systems when vowels,
perceptually close, are taken?
- what is the signification of the confusion when human
behaviour is considered?
Basically, CI and FFT lead to models of the language. These
models are based on the knowledge on speech, in perception and on acoustical
properties.
Acoustical analyses will be developed further in this text.
2.Material and methods
2.1 Cochlear implant use
Basically it can be considered that the acoustic wave undergoes
the transformations indicated in figure 1, from the acoustic wave to the
brain.
Brain Interpretation
|
Outer & Middle Ear
|
|
|
Inner
|
|
Auditory Pathways
|
|
|
|
|
|
|
Ear
|
|
|
|
|
|
|
|
Acoustical Signal
Figure 1: Classical stages in hearing.
Outer and middle ears transfer the acoustical vibrations to the
inner ear.
In the inner ear, vibrations are converted into electrical
stimuli which enter the auditory pathways.
The auditory pathways carry the influx to the brain, and some
transformations occur at this stage, mostly in the detection of basic features
(temporal and energy) in the signal.
In the brain, important and mysterious transformations take
place. Generally it is considered that the features are « printed »
in the temporal cortex, then matched with patterns already known by the
subject, and finally interpreted according to his background and knowledge of
the situation.
In the inner ear there is the organ of Corti which performs
the transformation from acoustical vibrations into electrical stimulus. When
this organ is totally non functioning (in both ears) the chain is broken,
leading to a deep cophosis, and the subject is completely deaf.
In order to beat this handicap, scientists have developed
cochlear implants intended to replace this deficient function.
The schematic structure of a cochlear implant is shown on figure
2. The description given is the DX 10 system of the French firm MXM, which was
used in this work.
Skin
Micro
Antenna 0
Speech processor
Antenna
Detection & distribution
Electrodes
Figure 2: The main two parts of a cochlear implant.
In the speech processor, the signal is sampled at a 15.6 kHz
rate, and a numerical analysis takes place. A FFT is performed and 64 spectrum
lins are calculated and grouped into 15 frequency bands ranging (table I) from
122 to 6558 hertz (Hz). Then 15 pulses with a fixed amplitude are constructed.
The duration of each pulse is proportional to the energy of the corresponding
frequency band. The pulses modulate a 3 MHz carrier which is transmitted,
through the skin, to the implanted part of the device.
After reception, the implanted electronic device demodulates
the carrier and distributes, sequentially, to 15 electrodes the electrical
energy contained in the pulses. According to their
place in the cochlea, the electrodes represent different
frequencies... and the brain must deal with them and «guess» the
content of the message formulated by the speaker.
|
Band
|
Frequency range
|
|
1
|
122-244
|
|
2
|
244-366
|
|
3
|
366-488
|
|
4
|
488-610
|
|
5
|
610-732
|
|
6
|
732-854
|
|
7
|
854-976
|
|
8
|
976-1098
|
|
9
|
1098-1342
|
|
10
|
1342-1708
|
|
11
|
1708-2196
|
|
12
|
2196-2806
|
|
13
|
2806-3660
|
|
14
|
3660-4880
|
|
15
|
4880-6588
|
Table I: Distribution of the frequency bands
(given in Hz) of the Digisonic DX-10 cochlear implant.
2.2 Acoustical material
Twelve French vowels were taken in the following context:
« c'est /v/ ça » (it is /v/ that)
where /v/ stands for the vowel. Vowels are written in phonetic
notation and indicated between slashes (table II). The use of
a context is intended to minimise the « side effects »
(coarticulation influence).
Two French speakers (male and female) participated in the
experiment. They were in their mid-twenties.
Bach-vowel was uttered 10 times leacling to 10 samples for each
clans:
|
Phonetic symbol
|
Typical word
|
Category
|
|
/a/
|
Pute (paw)
|
O
|
|
/a/
|
Pâte (pasta)
|
O
|
|
/à/
|
dans (in)
|
N
|
|
/e/
|
...té (summer)
|
O
|
|
/ce/
|
beurre (butter)
|
O
|
|
/8/
|
baie (bay)
|
0
|
|
/0/
|
le (the)
|
O
|
|
/i/
|
fille (girl)
|
O
|
|
lo/
|
porte (door)
|
O
|
|
/o/
|
beau (beautiful)
|
O
|
|
/I/
|
feu (lire)
|
O
|
|
/o/
|
bon (good)
|
N
|
|
/u/
|
sol (money)
|
O
|
|
/y/
|
brûler (to burn)
|
O
|
|
Ice I
|
brun
|
N
|
|
IÈ/
|
brin (blade)
|
N
|
Table II: French vowels used in this text; the category (Oral or
Nasal) is indicated in the right column
2.3 Signal recording
The diagram of the system used to record the signal is
indicated on figure 3. The acoustical signal was sampled at a 16 kHz rate by a
classical 16-bit sound-blaster card, using the corresponding routines in the
Windows package.
The microphone was a low-pass filter (cut-off frequency was 8
kHz) and it had also the antialiasing function. Then the signal was segmented
on disc files in order to select only the vowels to be studied.
In nasal phonemes, the air uses the nasal track. When phonemes
are oral, the nasal track is closed.
Windows routines
Sound blaster
^-> Personal computer
Disc
storage
Micro 0--
Figure 3: Block diagram indicating the recording of the phonetic
material.
2.4 FFT analysis
The FFT analysis was performed on data stored in the computer.
An overlap of 50% occurred between two consecutive windows. Samples were
weighted according to the Hamming formula. The duration of the analysis window
(frame) was 128 points corresponding to 8 milliseconds (ms) of signal.
Sixty-four spectrum fines were calculated and arranged according
to the 15 frequency bands of the Digisonic (table I).
On each utterance, 19 frames with a 50% overlap have been
taken in the middle of the vowel. Each frame led to a 15-dimensional vector,
where each component contained the energy of a frequency band. The final
values, for a band and for an utterance, were the average of the energies
calculated on the 19 frames.
2.5 Pulses recording
The Digisonic performs a spectrum analysis of the speech signal,
and the FFT parameters indicated in section 2.4 were taken for comparison
purposes.
A pulse, representing the energy, was associated to each band.
Pulses were automatically recorded using a special device
(Digistim) supplied by the manufacturer (figure 4).
The acoustical material to be analysed, previously recorded in
the computer, was played at the input of the Digistim (device supplied by MXM)
and the duration of the pulses (and the time
between two pulses) were detected and then stored into the
computer. This work was done under the control of a special software developed
by the manufacturer.
Sound Blaster
Desktop Computer
o
Cochlear Implant
Digistim MXM
Figure 4: Recording of the pulses.
2.6 Discriminant analysis
The discriminant analysis is a classical linear method of
classification [10]. Only two-class comparisons were performed in this study.
With the 16 classes (one class for each vowel), we had 16* 15/2 =120
comparisons.
Classically, the overlap between two classes is measured by X the
eigenvalue of the T-1E matrix, where:
T is the matrix of total covariance,
E is the covariance matrix of the centres of gravity of the
classes.
It can be proved [11] that X belongs to the [0,1] range in the
2-class case.
The largest eigenvalue was taken to assess the separation of the
classes. Let us remind that X =0 is a « perfect » confusion and X =1
is a « perfect » separation.
In order to make more sensible comparisons, the separation of
each pair of two classes has been also indicated, in projection on the main
eigenvector associated to the largest eigenvalue.
The results given by the
classical statistical formula (given below) have been shown:
- in
S'ab = a b
0. a2 cr
2b
where ma and mb are the means for each
class a and b and 6a2 and o the corresponding
variances.
3 Results and discussion
Results obtained with the female voice are given (tables III and
IV). Those obtained with the male voice are equivalent.
|
/a/
|
/a/
|
/a/
|
/o/
|
/e/
|
/s/
|
/I/
|
/i/
|
[É/
|
/o/
|
h/ /ce/
|
/6/
|
/u/
|
/y/
|
/êe' /
|
|
/a/
|
****
|
07.2
|
03.7
|
08.2
|
09.3
|
08.9
|
16.1
|
10.0
|
09.1
|
12.8
|
06.5
|
10.1
|
08.3
|
07.8
|
07.3
|
08.2
|
|
/a/
|
0.95
|
****
|
15.9
|
09.6
|
13.3
|
19.8
|
11.9
|
16.1
|
07.1
|
13.4
|
08.1
|
11.8
|
15.6
|
12.4
|
07.7
|
09.9
|
|
/à/
|
0.99
|
0.98
|
****
|
08.9
|
10.6
|
12.4
|
05.7
|
12.4
|
08.7
|
04.3
|
09.1
|
16.8
|
10.9
|
09.5
|
08.2
|
09.2
|
|
/o/
|
0.99
|
0.99
|
0.96
|
****
|
09.3
|
09.0
|
09.4
|
11.3
|
05.3
|
12.9
|
07.2
|
07.2
|
09.0
|
08.9
|
07.5
|
08.3
|
|
/e/
|
0.99
|
0.9
|
0.99
|
0.99
|
****
|
07.4
|
06.4
|
12.8
|
05.7
|
10.7
|
06.1
|
08.9
|
08.6
|
08.6
|
08.6
|
06.6
|
|
/d
|
0.99
|
0.99
|
0.97
|
0.99
|
0.99
|
****
|
05.3
|
10.1
|
09.0
|
07.1
|
12.8
|
08.4
|
07.7
|
07.3
|
15.8
|
07.2
|
|
/I/
|
0.99
|
1.00
|
0.98
|
0.97
|
1.00
|
0.96
|
****
|
09.7
|
11.7
|
05.4
|
08.3
|
07.5
|
06.6
|
06.3
|
08.6
|
06.3
|
|
/i/
|
1.00
|
1.00
|
0.99
|
0.99
|
0.99
|
1.00
|
0.99
|
****
|
14.3
|
09.0
|
13.8
|
08.6
|
10.6
|
05.3
|
12.0
|
03.7
|
|
re
|
0.96
|
0.97
|
0.99
|
0.98
|
0.99
|
0.99
|
0.99
|
0.99
|
****
|
12.2
|
06.7
|
11.1
|
12.3
|
12.6
|
05.1
|
11.0
|
|
/o/
|
1.00
|
0.99
|
0.97
|
0.99
|
0.98
|
0.99
|
0.97
|
1.00
|
0.99
|
****
|
08.8
|
07.6
|
05.6
|
05.4
|
08.4
|
06.2
|
|
/o/
|
0.99
|
0.99
|
0.98
|
0.96
|
0.99
|
0.97
|
0.96
|
0.99
|
0.99
|
1.00
|
****
|
08.0
|
11.4
|
10.3
|
05.4
|
09.0
|
|
/ce/
|
0.99
|
1.00
|
0.99
|
0.99
|
0.99
|
0.99
|
0.99
|
1.00
|
0.99
|
1.00
|
1.00
|
****
|
06.2
|
07.0
|
09.1
|
07.4
|
|
/6/
|
1.00
|
1.00
|
1.00
|
0.99
|
1.00
|
0.99
|
0.99
|
1.00
|
0.99
|
0.95
|
0.99
|
0.99
|
****
|
08.0
|
09.1
|
08.1
|
|
/u/
|
1.00
|
1.00
|
0.99
|
0.99
|
0.99
|
0.99
|
1.00
|
0.97
|
1.00
|
0.99
|
0.99
|
0.97
|
0.99
|
****
|
08.1
|
03.0
|
|
/y/
|
1.00
|
0.99
|
0.99
|
0.96
|
1.00
|
0.98
|
1.00
|
1.00
|
0.99
|
0.99
|
0.99
|
1.00
|
0.98
|
0.99
|
****
|
08.8
|
|
/Ce/
|
1.00
|
1.00
|
0.99
|
0.99
|
0.99
|
0.99
|
0.99
|
0.97
|
1.00
|
0.99
|
1.00
|
0.99
|
0.99
|
0.92
|
1.00
|
****
|
Table III : Separation between the vowels using the FFT; Xs are
on the bottom left and Sabs on the top right.
Results were calculated with the FFT model and with the cochlear
implant coding. 2,s are indicated on the bottom left and Sabs on the
top right of the tables. It can be seen that the vowels
were well separated in all the situations. Both models (FFT and
Implants) of vowel representations behaved equally.
Then, it was expected that some vowels would have a high
overlap. This is the case for /Ce/ and /V which are not easy to distinguish by
normal listeners in usual speech. In our example, the automatic separation is
almost perfect.
|
/a /
|
/a/
|
/à
|
/ /o / /e/
|
/s/
|
/I/
|
/i/
|
/"Ê / /o/ bo / /ce/ /ô/
|
/u/
|
/y/
|
/Ce/
|
|
/a /
|
****
|
15.0
|
08.4
|
13.9
|
10.9
|
19.7
|
06.7
|
18.6
|
13.6
|
06.8
|
10.9
|
17.6
|
14.7
|
12.2
|
20.1
|
14.8
|
|
/a/
|
1.00
|
****
|
14.9
|
05.4
|
29.2
|
15.8
|
11.7
|
29.4
|
07.9
|
24.0
|
09.3
|
11.7
|
33.4
|
26.1
|
24 6
|
08.8
|
|
lij /
|
1.00
|
0.99
|
****
|
14.2
|
14.4
|
12.1
|
07.1
|
22.2
|
08.7
|
05.1
|
16.0
|
10.9
|
07.5
|
13.0
|
17.4
|
08.1
|
|
/a /
|
0.99
|
0.98
|
0.99
|
****
|
20.5
|
10.2
|
11.2
|
19.9
|
05.2
|
28.1
|
12.5
|
05.7
|
19.6
|
18.2
|
17.3
|
06.1
|
|
/e/
|
0.99
|
1.00
|
0.99
|
1.00
|
****
|
13.2
|
13.3
|
11.0
|
11.2
|
09.2
|
15.5
|
11.3
|
04.0
|
06.5
|
09.3
|
20.5
|
|
/s/
|
0.99
|
0.99
|
0.99
|
0.99
|
1.00
|
****
|
06.8
|
18.3
|
07.2
|
18.1
|
15.4
|
10.7
|
15.4
|
13.8
|
11.8
|
18.8
|
|
/V
|
0.99
|
0.99
|
0.99
|
0.99
|
0.99
|
0.96
|
****
|
22.0
|
06.8
|
04.2
|
08.7
|
08.3
|
16.7
|
14.7
|
17.3
|
08.3
|
|
/i/
|
1.00
|
1.00
|
1.00
|
1.00
|
0.97
|
1.00
|
1.00
|
****
|
22.3
|
19.8
|
23.9
|
11.8
|
13.1
|
05.8
|
07.2
|
25.4
|
|
ft' /
|
1.00
|
0.99
|
0.99
|
0.95
|
1.00
|
0.98
|
0.97
|
1.00
|
****
|
10.2
|
07.0
|
04.2
|
12.8
|
17.6
|
18.0
|
04.7
|
|
/o/
|
1.00
|
1.00
|
0.97
|
1.00
|
1.00
|
0.99
|
0.96
|
1.00
|
0.97
|
****
|
09.6
|
15.3
|
07.2
|
11.6
|
19.3
|
13.8
|
|
bo /
|
0.99
|
0.98
|
0.99
|
0.99
|
1.00
|
0.99
|
0.99
|
1.00
|
0.97
|
0.98
|
****
|
11.4
|
14.0
|
17.4
|
19.0
|
05.2
|
|
le /
|
1.00
|
0.99
|
0.99
|
0.98
|
0.99
|
0.99
|
0.98
|
0.99
|
0.97
|
0.98
|
0.98
|
****
|
11.7
|
11.1
|
09.9
|
08.4
|
|
/5/
|
0.99
|
1.00
|
1.00
|
1.00
|
0.97
|
1.00
|
1.00
|
0.97
|
1.00
|
0.99
|
1.00
|
1.00
|
****
|
07.5
|
12.0
|
18.5
|
|
/u/
|
0.98
|
1.00
|
0.99
|
0.99
|
0.98
|
0.99
|
1.00
|
0.95
|
1.00
|
0.99
|
0.99
|
0.99
|
0.95
|
****
|
05.4
|
17.3
|
|
/y/
|
1.00
|
1.00
|
0.99
|
1.00
|
0.98
|
0.98
|
0.99
|
0.94
|
1.00
|
1.00
|
0.98
|
0.98
|
0.98
|
0.95
|
****
|
16.0
|
|
/Ce/
|
1.00
|
1.00
|
0.99
|
1.00
|
0.99
|
0.99
|
0.99
|
1.00
|
0.93
|
0.99
|
0.94
|
0.99
|
1.00
|
1.00
|
0.99
|
****
|
Table IV: Separation using the cochlear implant. Considering
these results two comments can be made:
1) The vowels were carefully spoken in good signal to noise ratio
conditions (soundproof room); it would be interesting to see the results in
more « natural » (noisy) situations.
Comparisons were made on the voice of one speaker only... in the
case of several speakers, it would be useful to reconsider the situation.
2) The representation which was taken (vectors and eigenvalues in
a discriminant analysis)
did not appear to be a good model of human behaviour with the
vowels. People are more used to cope with the variability which is included in
normal speech conditions.
Further studies are needed, in order to investigate deeper
that matter. For instance, a look at the structure of the representative
vectors cannot be avoided. A model of human behaviour in such conditions is
still to be made and the spectral distribution presented in this work must be
improved to match better human psycho-acoustical performances.
IV Conclusion
Vowels representation by a cochlear implant, using the model of a
spectrum vector and a discriminant analysis (or a statistical distance), is
very efficient to separate the vowels.
This efficiency was surprisingly high because several vowels
(acoustical objects) which are similar perceptually were perfectly
separated.
Results given by a FFT analysis were very similar to those
obtained with a cochlear implant.
This work indicates that further studies are needed to set up a
model more adapted to human behaviour.
Acknowledgements
The authors acknowledge the participation in this experiment of
the students involved in the project: M. Doutiaux, S. Faucher, B.
Fuselier, E. Laribe, D. Rousseau.
References
[1] Fant R. « Auditory patterns of speech »,
W. Dunn edit., Cambridge, M.I.T. Press, (1967), 111-125.
[2] Roberts L. « Speech and brain mechanisms »,
Princeton Univ. Press, (1959).
[3] Berger-Vachon C., Gallego S., Bera J.C.,.Arnoux E., Vissac
C. «A Model of vowels representation using a cochlear implant »,
Advances in Intelligent Systems, 526-532 Ed. F.C. Morabito, I. 0 . S .
Press, Amsterdam , (1997)
[4] Perkell J., Klatt D.H. « Invariance and
invariability in speech processes », Lawrence Earlbaum, Hullsdale
N.J., USA, (1983).
[5] Serniclaes W., De Guchteneere R., Secqueville T.,
Bachelot G., Genin J., Meyer B., Chouard C.H. « Objective evaluation
of vowel identification with the Digisonic cochlear implant »,
Audiology, 35, 23-36 (1996).
[6] Belaieff M., Dubus P., Leveau J.M., Repetto J.C., Vincent
P. « Sound processing and stimulation coding of Digisonic DX-10
15-channel implant », Advances in cochlear implantation, Ed I.J.
Hochmair (Innsbruck), 198-203 (1996).
[7] Robinson K., Summerfield A. « Adult auditory
learning and training », Ear & Hearing, 51S- 65S (1996).
[8] Dorman M.F., Loizou P.C. « The identification of
consonants and vowels by cochlear implant patients », Ear &
Hearing, 162-166 (1998).
[9] Skimmer M.W. et al « Speech recognition at
stimulated soft, conversational and raised to loud vocal efforts by adults with
cochlear implants », J. Acoust. Soc. Am., 101, 3766-3782 (1997).
[10] Sebestyen G.S. « Decision making processes in
Pattern recognition », Mc Millan (1962).
[11] Lebard L., Morineau A., Fénelon J.P. «
Traitement des données statistiques », Dunod-Paris
(1982).
b/ Restitution acoustique des signaux provenant de
l'électrodogramme
Précédemment, nous avons montré que
l'information résultant de l'implant cochléaire
Digisonic® est très pertinente. La question maintenant
posée est de savoir si la manière dont l'information est
transmise au système auditif nécessite un recodage. En effet,
l'information, très pertinente pour l'ordinateur, peut être mal
décryptée par les voies auditives.
Il nous a paru intéressant de simuler le signal entendu
par les sujets implantés cochléaires, par une restitution
acoustique des électrodogrammes, afin d'évaluer son
intelligibilité lors d'un traitement normal via les voies auditives.
| 
