Building and sharing multilingual speech resources,
using ERIM generic platforms
Georges FAFIOTTE
GETA, CLIPS, IMAG-campus (UJF, Grenoble 1 Univ.)
385 rue de la Bibliothèque, BP 53
F-38041 Grenoble cedex 9
France
georges.fafiotte@imag.fr
Abstract
In the framework of projects ChinFaDial and
ERIM we have developed in recent years
several platforms allowing to handle various
aspects of bilingual spoken dialogues on the
web —mainly, spontaneous speech corpus
collection through distant human interpreting.
Current development of the core ERIM-Interp
and ERIM-Collect platforms now includes
multimodal user interaction, integration of
some machine aids (such as speech turn logs
through speech recognition, or tentatively
speech machine translation, both based on
server-grounded market products), and next,
online aids to speakers and/or interpreters.
First collected data should be made available
on the web in fall 2004 (DistribDial) along
with, as soon as available, a robust version of
the collecting platform, in order to promote
collaborative building, and sharing, of "raw"
unannotated multilingual speech corpora.
A variant of the ERIM environment is to
extend to distant e-training in interpreting,
possibly creating situations which should in
turn, in our view, foster larger-scale data
collection and sharing in open access mode.
Keywords
Bilingual speech corpora, collaborative corpus
collection, spontaneous dialogues, Web-based
interpreting, multilingual communication,
open-access resources, resource mutualization.
Introduction
Ongoing burst in the development of both portable
telecommunications tools open to Internet
transactions, and videoconferencing means, is
creating rapid expansion of teleservicing and
telebusiness applications with spontaneous
dialogue, information inquiry, distant negotiation,
etc. Multilingualism, now in spoken transaction as
it has been in written one, appears as a key issue in
distant communication, with sensitive questions,
both in supporting the diversity of the native or
origin language of conversing users (particularly
within the opening European economic area), and
in bringing some kind of balance between main
"linguae francae" (common languages). Thus new
stakes arise in enhancing distant web-based on-line
interpreting services.
Meanwhile, Speech Machine Translation (SMT)
steadily takes steps towards style spontaneity and
multilingualism. In this context though, we face a
notorious lack of large open-access corpora of
bilingual spoken dialogues.
This led us to study, to model and propose a set of
generic platforms, aiming at enhancing distant
multilingual multimodal oral communication with
full recording and collecting facilities, also
addressing expectations from the MT systems
engineering community.
The paper first looks over project motivation, then
introduces the interpreting and collecting platforms
presently available in the ERIM family, with
current variants. It then reports on their first use in
collecting domain-oriented spontaneously spoken
French-Chinese dialogues. Finally we present
ongoing or planned development, advocating for
collaborative building and voluntary sharing of
resulting multilingual resources.
1. Motivations, early prototyping
1.1 Developing multilingual linguistic
resources
It is widely recognized that realistic and large
corpora are key resources for building Speech
Recognition (SR) and Speech MT systems. If the
Web has recently been put to use as the largest
possible corpus, modeling casual spontaneous
spoken language requires transcribed speech
corpora of hundreds of hours.
Speech translation systems thus need large parallel
translation corpora of transcribed and aligned
spontaneous utterances in dialogue context, ideally
with complete sets of parse trees. However, few
such corpora have been developed (by NEC, ATR
and a few others), and these are not publicly
available. Why not? Because these corpora are
very expensive to transcribe once collected, and to
annotate. After so much time has been spent in
compiling a corpus, giving it away seems
unreasonable.
Besides, a future research objective is to use
collected corpora for studying and modeling real
life spontaneous spoken language and dialogues,
and possibly to investigate if and how specific
linguistic traits can be expected depending on
specific dialogue situations, translation process
settings, or various multimodal interaction means.
For instance, two speakers in a bilingual dialogue
may hear one another's original speech or not, they
may use video or fixed images, etc. Their linguistic
behavior is expected to vary accordingly: the
number of clarification sub-dialogues may vary;
third person use or indirect speech may be used
more in the presence of a speech translation system
than with a human interpreter; the use of deictic
and anaphoric elements may turn out to depend on
the use of visible markable objects on whiteboards,
maps, images.
With these considerations in mind, we thus
endeavoured to propose open-acces corpus
resources —and therefore open-access collecting
resources—, in order to ease collaborative building
of "raw" unannotated multilingual translated
speech corpora, likely taking advantage of new
web-based interpreting situations or scenarios.
1.2 Enhancing multilingual communication
on the Web
Some companies have already developed
proprietary network-oriented interpreter's cubicles,
which are the counterparts of existing fixed
installations for interpreting in multilingual
meetings (for example at the UN or EU). However,
the associated code is not available for research.
Furthermore, our typical scenario is somewhat
different from that of classical interpreting, where
interpreters are available for the entire duration of
the conversations. We rather allow two situations:
• "conference call": speakers establish a schedule,
and book a time slot with an interpreter,
• "on demand interpretation": interlocutors try to
converse using whatever knowledge they may
have of their interlocutor's language, or of a
third common language. When the language
barrier impedes communication, they ask an
available interpreter to jump in to help.
Apart from these practical motivations, we also
wish to conduct experimental studies on the effect
of combining multimodal resources on bilingual or
multilingual conversations. Thus, full recording
facilities were required anyhow.
1.3 Pre-ERIM platforms
Other studies of human "consecutive"
interpretation have employed multimodal Wizard
of Oz platforms (e.g. the EMMI plateform, that we
experienced at ATR-ITL for bilingual pilot-
experiments [Fafiotte & Boitet, 1994] [Loken-Kim
& al., 1994]), or monolingual multi-Wizard
architectures have been modelled in a multimodal
setting (NEIMO [Coutaz & al., 1996]). Thus our
first objective was to produce a simulator of
automatic speech translation systems in the same
spirit, to gain experience and collect data.
We first built prototypes of a Speech MT Wizard
of Oz simulator, Sim* [Fafiotte & Zhai, 1999] (to
be read as "Sim-Star", since being a parallel
platform to the C-STAR II CLIPS environment).
They were designed to run on the Internet, and
were originally used on the intranet of CLIPS-
GETA. Network-based communications were
handled by a client-server communication module
developed in Tcl/Tk. Participants could see and
hear each other and share an electronic whiteboard,
using MBone resources.
The idea of using Wizard of Oz techniques in this
context proved quite impractical, and thus was
abandoned. Even if an acoustic filter was used to
deform the interpreter’s voice, participants
perceived that a human was speaking. In the end,
we realized that, even for true automatic high
quality interpretation, there actually might well be
a real human "warm body" in the loop anyway.
Thus a realistic design for online interpretation
could integrate both human and machine
interpretation for "partially automatic" Speech MT.
The successive ERIM platforms have been
implemented on this basis, in parallel at CLIPS
with integrating the French language into
multilingual Speech Machine Translation within
C-STAR and NESPOLE! international projects.
ERIM stands in French for Network-based
Environment for Multimodal Interpreting.
2. Distant human interpreting, as a collecting
scheme for multilingual spoken dialogues
2.1 Context
At CLIPS-GETA, one of the ultimate research
goals in Speech MT is to build systems for
automatic or partially automatic Speech
Interpretation (i.e. "synergic" user-aided translation
of speech). Much progress has been made in this
area over the past twelve years. NEC produced the
first speech translation demo in September 1992,
within the tourist domain, but the most widely
known coordinated research efforts to date include
the C-STAR projects (now a 7-language
international Consortium for Speech Translation
Advanced Research) [http://www.c-star.org], the
European NESPOLE! project [http://nespole.itc.it],
the German Verbmobil [http://verbmobil.dfki.de]
project, the US DARPA Communicator program
with the Galaxy Communicator Software
Infrastructure [http://fofoca.mitre.org/doc.html]
[http://www.darpa.mil/ito/research/com/index.html]
[http://www.sls.lcs.mit.edu/sls/whatwedo/architecture.html].
All have demonstrated platforms enhancing
spontaneous speech processing in multilingual
person-person or person-system communication,
always in restricted domains. CLIPS is firmly
involved in this action, while being in charge for
integrating the French language in the C-STAR
and NESPOLE! environments.
At the same time, we strongly believe that human
interpreters will remain vital, both as irreplaceable
suppliers of relevant nuances and as models for
automatic or partially automatic systems.
Human interpreting, too, will inevitably be carried
out through the Web and its raising applications.
Thus we foresee a continuing need for research on
Web-based interpreting, and for data collection of
realistic general-purpose or domain-oriented Web-
based interpreting sessions.
2.2 Functionals of the ERIM human
Interpreting platform
The ERIM-Interp network-based environment
consists of a central communication server, two
speaker stations, one interpreter station (cf. Fig. 1),
with a multimodality server (exchange of short
typed messages, whiteboard with shared pictures
or files, and shared pointing and marking). To
avoid complex problems due to turn overlap, we
have adopted a push-to-talk discipline up to now.
The current implementation of ERIM-Interp, in
Tcl/Tk, is platform independent (and runs on
Windows, MacOS, eventually Linux), and uses an
adapted version of the CommSwitch written by
CMU for the CSTAR-II project.
It is also flexible: the CommServer can be hosted
on a dedicated station or on any user workstation,
two speakers may share the same station (in a
"visit" situation), the scenario can be extended to
include more than two interlocutors, more than one
interpreter (in "one-way" interpreting situations),
and hence possibly more than two languages.
3. Bilingual spontaneous speech collection
3.1 As the next step taken then, the ERIM
Collecting platform
We have then developed the ERIM-Collect variant,
intended to collect corpora (cf. Fig. 1), moreover to
enhance collaborative generation and use of
bilingual speech corpora; namely to:
• collect only "raw" data (web-based spontaneous
dialogues in any language pairs), as multimodal
as possible —with no built-in annotation
scheme intended yet,
• motivate volunteers to produce the data,
• induce volunteering by offering free service (on
one of the ERIM variants described here), in
exchange for free data (users should agree to
"donate their speech to science"),
• distribute the data as freeware (via GPL
licensing) on the Web, in a "re-playable" form:
for each dialogue, descriptors indicate essential
(anonymous) facts about the participants, along
with the list of turns, indications of files,
speakers, and time stamps for each turn,
• make it possible for other researchers to enrich
the corpora by adding annotations in parallel
files, again sharable through the web; they
might use an extended version of the "Replay"
facility (cf. Fig. 3), with consensus on a shared
file structure and XML descriptors format,
• develop the collection platform so that it can
itself be offered as freeware on the Web.
Accordingly, ERIM-Collect (currently 350 Kbytes
of code in Tcl/Tk) was defined as an extension of
ERIM-Interp:
• ERIM-Collect is language-independent,
• data is recorded locally during the dialogue;
speech files are in PCM 22kHz-16bit-mono
format,
• session and speech turns descriptor files are
now in XML format,
• after the conversation, local descriptors and
files are transferred then structured in corpus
bases on a Collection Server,
• everything possible should be recorded: speech,
short texts, whiteboard events, video, objects
which the speakers refer to (e.g. file names and
urls). In the current version 3 of ERIM-Collect,
voice and short texts are collected; whiteboard
actions and video are currently added.
Speaker 1
CORPUS
 translation
 translation
into French
French
turn
Communication
Server
+
Interpreter
Speaker 2
Chinese
turn
translation
into Chinese
CORPUS
CORPUS
 translation
 WhiteboardWhiteboard
2a
4a
3
1a
5a
5b
1b
2b
4b
6
Figure 1: ERIM-Interp / ERIM-Collect
We describe here (cf. Fig. 1) a basic exchange
within a French-Chinese collection session. First
(1), the French interlocutor takes a turn of one or
more utterances. This turn (speech, descriptors) is
recorded locally (1a), and transmitted (1b) to the
Interpreter and the CommServer which broadcasts
it across the virtual room established for the
conversation. The interpreter listens to the turn and
(2) translates it into Chinese. The translated turn is
recorded locally (2a) and broadcast (2b). The
Chinese participant listens to the translation (3)
and then answers (4). Again, his answer is stored
locally and broadcast (4a and 4b). The interpreter
then translates it into French (5) and the translation
is stored locally (5a) and broadcast (5b).
In order to create various experimental settings, we
may unlock the reception of some messages for
some participants. For instance in (1b) the French
voice could be made audible for the Chinese
participant.
Figure 2 shows the screen which is presented to a
conversational partner, as presently prototyped for
the ERIM-Collect platform.
Figure 2: Speaker's screen
As for playback of a
previously recorded
bilingual dialogue, a
full reconstruction is
available. Simplified
visual tracking is
provided as shown in
Figure 3. One can
extract monolingual
versions of the
dialogues.
A first version of the
DistribDial / Replay
component (and web
site) for such replays
has just been
completed.
Figure 3: Playback of client, interpreter, and agent utterances
Successive versions of ERIM-Collect have been
used for collecting first domain-oriented
spontaneous speech corpora (hotel reservation) in
Grenoble and Beijing (cf. 4.2).
3.2 Providing online aid to interpreters
and/or speakers
In our "on demand interpretation” scenario,
interpreters may be asked to jump from one
conversation to another, and thus from one topic to
another. This conversation switching is likely to be
quite difficult, and stressful. Thus machine aids
could be welcome: communication aids and
language aids. We also envisage providing
machine aids for the conversational partners, to
help them do without interpreters so far as
possible, if necessary.
The currently implemented "communication aids"
include facilities to
• see and hear others (participants and
interpreters),
• share data, possibly modifiable, markable, and
"pointable" through the whiteboard,
• access an agenda for scheduling rendezvous.
Possible "language aids", to both the human
interpreter and the speakers, are of three kinds:
• access to dictionaries via typed or voiced
requests, and via automatic word spotting
followed by filtering, dictionary look-up, and
presentation in a dedicated window,
• speech recognition, to alleviate difficulties of
oral understanding when not using the
interpreter, and to produce a log of the
conversation (which can additionally help an
interpreter jump in), after possible reduction,
• fully or partially automatic speech translation.
At this time most communication aids have been
implemented. The scheduling agenda is global for
an ERIM site, but each user handles it through a
personalized view (cf. Fig. 4).
Figure 4: Window of user agenda
Language aids are the next step. An interface to
existing free dictionary resources on the Papillon
site [http://www.papillon-dictionary.org] should be
added soon. A speech recognizer has been
connected to the platform in another ERIM variant
(the automatic interpretation pilot setup ERIM-
paST). This Speech-To-Text facility could help as
well to issue some draft transcripts during the
dialogue.
3.3 Adding partially automatic Speech MT
An ERIM-paST (partially automated Speech
Translation) platform is in progress at CLIPS in
Grenoble, originally in cooperation with Spoken
Translation Inc. (Berkeley). It aims at eventually
providing some languge aids to speakers who
"converse by themselves", and at allowing data
recording of partially automatic interpreted
dialogues (as a testing ground for Speech MT
systems development, testing or tuning, at CLIPS).
Experimentation with interactive disambiguation
methods derived from the LIDIA project [Boitet &
Blanchon, 1994] is also expected.
The detailed description of this ERIM variant is
beyond the scope of this paper. Briefly stated, the
goal is here a generic modular integration, through
plug-in, of Speech MT modules (speech
recognizers, text-to-text translators, speech
synthesizers), either research components (for their
fine testing and tuning) or off-the-shelf products.
Objective is to carry out comparative assessment
of their results, or possibly contrastive evaluation
with the human production of an interpreter "warm
body".
A first version of ERIM-paST is currently being
prototyped, while integrating server-based (Philips,
Linguatec, Scansoft) market components.
4. First corpus collection, towards a
collaborative building/sharing scheme
4.1 Platform assessment: distant collection
Distant collection is also being tested, but in our
first experiments Voice/IP still proved problematic
when two turns overlapped. New efficient basic
software and connection improvements are under
evaluation. Record-then-send or record-while-
sending (streaming) modes are available.
We may retain facilities for transmitting sound
through phone lines. These might be used in
operational contexts by telephone operators, such
as Prosodie in France: since this company is also
an Internet service provider, it can merge both
tracks into a single communication.
Distant connection data is summarized in Figure 5.
Experiments
(Grades
from 0 to 5)
text
voice:
record
then
send
voice:
record &
send
(streaming)
voice: same
with
overlapping
Streaming — — + +
Connexion:
Internet
100
Mbit
= = =
Reception
quality
5 5 3 1
Speed of
exchange
5 2 4 5
Reliability 5 5 4 1
Special
problems /
phenomena
None
User
wary
(too
slow)
Some
micro-cuts,
but good
overall
quality
Unusable,
bandwidth
too large
Figure 5:  Oral communication over the web
4.2 The ChinFaDial project, French-Chinese
speech corpora
The system has been used in the ChinFaDial
project for collecting bilingual French-Chinese
interpreted spontaneous spoken dialogues, in the
hotel reservation domain. This 3-year project was
funded by LIAMA, a joint French-Chinese
laboratory under both French INRIA and Chinese
CAS and MOST supervision. Our partner is the
Chinese Information Processing group at NLPR
(National Laboratory for Pattern Recognition), a
research team within the Institute of Automation,
Chinese Academy of Sciences (CAS-IA).
In ChinFaDial we have used intranets in Grenoble
or in Beijing, with 3 participants using headsets,
located in one or in 2 different buildings. It was
possible for 2 speakers to share the same
workstation, but we have mainly used the regular
3-station setting for the French-Chinese data
collection. Some 10 hours of spontaneous
translated spoken dialogues on "hotel information
and reservation" have been recorded thus far. They
produce about 43kBits per second.
Figure 5 shows a dialogue fragment transcription.
We do not plan currently to transcribe or annotate
corpora, but others will be very welcome to do so.
Participants to this first data collection have been
at this time:
Chinese French Total
Fr-Ch
Interpreters
2 2 4
Interlocutors 3 3 6
There are 65 recorded dialogues with these
characteristics:
Minimum Average Maximum
Duration (sec) 457 635 874
Number of turns 28 52 78
Turn length (sec) 4 12 57
Figure 5: Dialogue between a French hotel manager and a Chinese client (manual transcript)
4.3 Ongoing developments, to promote
collaborative corpus building
A website with a small ‘DistribDial’ server has
been prototyped to freely distribute the sound files
and their descriptors, and a Replay module. Our
goal is to extend it to allow other groups to
contribute to the site whatever annotations they
may have created, and to share them under the
same conditions (GPL). They should only agree to
share a common file base structure and a flexible
XML desciptor format for each annotation file.
Corpus collection in French-Chinese will extend.
Further data collection using ERIM-Collect just
started (spontaneous dialogues in French and
Vietnamese, Tamil, Hindi), under support of AUF
(University Agency for French-Speaking
Communities), within the VTH-Fra.Dial project.
We are also considering distributing an ERIM-
Collect "hardened" version on DistribDial, after
strengthening robustness and usability, so that
others can use it to do their own spoken dialogue
collection.
4.4 Planned e-Training extensions: use of the
platform to involve volunteer interpreters
Data collection being time-consuming all the same,
our goal is not to do too much of it for its own
sake, but to get it as byproduct of some
"mutualized" use of the platform, in the open
access mode.
Professional interpreters are unlikely to help on a
non-profit basis, since interpreting is their
livelihood. Improving junior interpreters or even
advanced student interpreters, however, may find
Web-based cooperation to be a good way of
learning or perfecting their trade in real life
situations.
We aim to induce volunteer interpreters or students
of interpretation to translate bilingual dialogues
online, by exchanging this on-line help for free use
of our Web-based lab for e-training in
interpretation.
We plan to develop an ERIM-Training variant
platform, an e-training extension, with full
recording of all speech interaction and any
multimodal event. Actually we already simulated
the functional architecture of it, using the current
ERIM-Collect in a multi-interpreter setting.
Different scenarios and settings can be envisaged.
For example, in a distant training or practice
situation, for a student interpreter: the student
might be alone, gaining experience, or might be
with an instructor, who could supervise or take
over.
At the 2008 Olympic Games in Beijing, as another
example, good student interpreters could be asked
to aid bilingual communication in exchange for
academic credit, and free tickets. Assume, for
instance, that a French speaker and a Chinese
speaker want to converse. They could then go to a
PC, activate ERIM-Interp with ERIM-Assist for
French-Chinese, click on the icon of an available
interpreter, and begin a mediated conversation,
which would be recorded if participants agree
while using the service free of charge.
4.5 Building and sharing multilingual speech
resources
We advocate and expect ERIM-Collect, once
proposed in an open-access mode on the Web, to
be willingly and freely operated by other
researchers, under an agreed collaborative
framework to be set up, with minimal method and
technical consent on collecting procedures and
corpus characteristic profiles, in order to bring
building and sharing of raw multilingual speech
corpora to a more rapid expansion.
Collaborative annotation work could take place as
well, again with simple agreed procedures on
content and descriptor files formats, and on a
public use scheme.
Such tools, and their open use, could as well
underlie valuable action towards supportive
protection of "smaller languages", among others
minor European languages, while for instance
fostering distant learning of interpreting, and while
easing the use of low-cost or even free interpreting
facilities over the net.
5. Unification of ERIM platform variants
Work is now beginning on the integration of the
different platforms presented here into one single
multifunctional ERIMM system [Fafiotte & Boitet,
2003], for enhancing free multilingual multimodal
network-based communication with distant
interpreting and corpus collection.
Numerous technical issues arise in this effort. For
instance, it is not immediately clear how the
CommServer will accommodate server-based
interactive lexical disambiguation during
translation; or how to secure efficient streaming
data transmission in a multicast scheme. Even so,
the platform independence and plug-and-play
generic architecture of ERIM set components make
this integration effort quite realistic, in spite of the
number and diversity of functions to be integrated.
Conclusion
We have presented several platforms developed in
the long-range ERIM project. Each platform can
aid in the study of spontaneous cross-lingual
communication on the Web. The core platform is
ERIM-Interp for Web-based human interpretation.
ERIM-Collect is a deliberate development of the
latter, dedicated to multilingual "raw" speech
corpus building, and intended to alleviate the
current scarcity of data —particularly open data—,
and which can also support the construction of
speech translation systems.
ERIM-Assist will add various machine aids for
interpreters and conversational partners, while
ERIM-paST (only briefly mentioned here) includes
components for partially automatic speech
translation.
We then reported on a first collection of
spontaneous bilingual interpreted spoken dialogues
for French-Chinese. This data, along with the
collecting framework itself, will be distributed in
the near future on the Web as shareware or GPL-
ware, under a DistribDial component.
We are looking for funding to create ERIM-
Training —a further extension of ERIM-Interp—
which could serve as a valuable "Web-based
language lab for interpreting" for distant e-training,
while also providing new facilities for language
learning.
We plan to continue research in the ERIM
framework by collecting and distributing more data
concerning more languages (Vietnamese, Tamil,
Hindi to French). Data collection should be
enhanced by a unified version of ERIM, offering
all the functionalities of the platform variants.
More specifically, we hope that junior interpreters
or advanced students in interpreting will volunteer
to interpret and to practice with ERIM-Training,
while users would agree to give their dialogues to
science in exchange of using ERIM-Interp for free.
Acknowledgements
This work has been supported by CLIPS-IMAG
(UJF University Grenoble 1, CNRS, INPG) and
funded in part by the LIAMA French-Chinese
Laboratory (ChinFaDial project), and by the
Rhône-Alpes Region (ERIM project). Corpus
collecting action is currently supported by AUF-
LTT (University Agency for French-Speaking
Communities, VTH-Fra.Dial project).
Our thanks go to Zhai JianShe (Nanjing
University, China) for early prototyping, to Julien
Lamboley (at INSA, Lyon, France) for platform
development, to members of the GETA and
NLPR/CASIA-Beijing teams and to Brigitte
Meillon at CLIPS-MultiCom, for their
participation in data collection and related
experiments.

References
Boitet C. & Blanchon H., 1994. Multilingual
Dialogue-Based MT for Monolingual Authors:
the LIDIA Project and a First Mockup. Machine
Translation 9/2/94, pp. 99-132.
Coutaz J., Salber D., Carraux E. & Portolan N.,
1996. NEIMO, a Multiwork station Usability Lab
for Observing and Analyzing Multimodal
Interaction. Proc. CHI'96 companion.
Fafiotte G. & Boitet C., 1994. Report on first
EMMI Experiments for the MIDDIM project in
the context of Interpreting Telecommunications.
MIDDIM report TR-IT-0074 GETA-IMAG &
ATR-ITL, Aug. 94, 11 p.
Fafiotte G. & Boitet C., 2003. ERIMM, a platform
for supporting and collecting multimodal
spontaneous bilingual dialogues. IEEE NLP-
KE2003, Beijing, 26-29/10/03, 6 p.
Fafiotte G. & Zhai J.-S., 1999. A Network-based
Simulator for Speech Translation. Proc.
NPLRS’99, Beijing, 5-7/11/99, B. Yuan, T.
Huang & X. Tang ed., pp. 511-514.
Furuse O., Sobashima Y., Takezama T. & Uratani
N., 1994. Bilingual corpus for speech
translation. Proc. AAAI-94 Workshop on
Integration of Natural Language and Speech
Processing, Seattle, Washington, USA, 31/7-
1/8/94, ATR Interpreting Telecommunications.
Loken-Kim K.-H., Yato F. & Morimoto T., 1994.
A Simulation Environment for Multimodal
Interpreting Telecommunications. Proc. IPSJ-
AV workshop, March 94, 5 p.

C-STAR.         http://www.c-star.org

DARPA sites.
http://www.darpa.mil/ito/research/com/index.html,

DARPA sites.
http://fofoca.mitre.org/doc.html

GALAXY system architecture site.
http://www.sls.lcs.mit.edu/sls/whatwedo/ architecture.html

site web NESPOLE!        http://nespole.itc.it

site web PAPILLON.
                      http://www.papillon-dictionary.org

VERBMOBIL site. http://verbmobil.dfki.de
