Providing Accessibility to Hearing-disabled by a Basque to Sign
Language Translation System
María del Puy Carretero, Miren Urteaga, Aitor Ardanza,
Mikel Eizagirre,
Sara García and David Oyarzun
Vicomtech-IK4 Research Center, P. Mikeletegi, 57, 20009 San Sebastián, Spain
Keywords: Virtual Character, Automatic Translation, Sign Language, LSE, Natural Language Processing.
Abstract: Translation between spoken languages and Sign Languages is especially weak regarding minority
languages; hence, audiovisual material in these languages is usually out of reach for people with a hearing
impairment. This paper presents a domain-specific Basque text to Spanish Sign Language (LSE) translation
system. It has a modular architecture with (1) a text-to-Sign Language translation module using a Rule-
Based translation approach, (2) a gesture capture system combining two motion capture system to create an
internal (3) sign dictionary, (4) an animation engine and a (5) rendering module. The result of the translation
is performed by a virtual interpreter that executes the concatenation of the signs according to the
grammatical rules in LSE; for a better LSE interpretation, its face and body expressions change according to
the emotion to be expressed. A first prototype has been tested by LSE experts with preliminary satisfactory
results.
1 INTRODUCTION
This paper presents a modular platform to translate
Basque into LSE. The modularity of the platform
allows the input to be audio or text depending on the
needs and the available technology of each
application case. The main objective of translating
spoken language into Sign Language is to allow
people with hearing loss to access the same
information as people without disabilities. However,
beside social reasons, this research project has also
been motivated by legal and linguistic reasons
1.1 The Importance of Sign Languages
Sign Languages are the natural languages that deaf
people use to communicate with others, especially
among them. The following facts on Sign Languages
and deaf people highlight the need of doing research
on Sign Languages:
Not all deaf people can read. A deaf person
can read well if deafness has come at adult
life. However, in most deafness cases at
prenatal or infant age the ability to interpret
written natural text does not develop, due to
the difficulty to acquire grammatical and void
word concepts. Therefore, these types of deaf
people are unable to read texts and/or
communicate with others by writing.
Lipreading. Some deaf people can read lips,
but it is not a general ability. Furthermore,
lipreading alone cannot sufficiently support
speech development because visual
information is far more ambiguous than
auditory information. Sign Language involves
both hands and body. Hand gestures should
always be accompanied by facial and corporal
expressiveness. The meaning of a sign can
change greatly depending on the face and
body expression, to the extent that it can
disambiguate between two concepts with the
same hand-gesture.
Sign Language is not Universal. Each country
has its own Sign Language, and more than one
may also co-exist. For example, in Spain both
Spanish Sign Language and Catalan Sign
Language are used. A deaf person considers
his/her mother tongue the sign language used
in his/her country.
1.2 Legal Framework
According to statistics collected by the National
Confederation of Deaf People in Spain (CNSE),
256
Carretero M., Urteaga M., Ardanza A., Eizagirre M., García S. and Oyarzun D..
Providing Accessibility to Hearing-disabled by a Basque to Sign Language Translation System.
DOI: 10.5220/0004748502560263
In Proceedings of the 6th International Conference on Agents and Artificial Intelligence (ICAART-2014), pages 256-263
ISBN: 978-989-758-015-4
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
there are approximately 120.000 deaf people in
Spain who use Sign Language as their first language.
Backing up deaf people's rights, the Spanish General
Law of Audiovisual Communications (CESyA,
2010) lays down that private and public channels
have to broadcast a certain amount of hours per
week with accessible contents for disabled people.
Table 1 summarizes the amount of hours per week
that broadcasters must provide with accessible
content for deaf people by the end 2013:
Table 1: Accessibility requirements of the Spanish
General Law of Audiovisual Communication.
Private service Public service
Subtitles 75% 90%
Sign language 2 hours 10 hours
Audio description 2 hours 10 hours
In order to adapt the contents to Sign Language
an interpreter is necessary. Adapting contents to
Sign Language is expensive and requires several
hours of work for off-line contents and more than
one interpreter for live contents.
In this paper we present a platform to adapt not
only audiovisual contents but also other kind of
information to Sign Language by using virtual
characters whose role is to interpret the meaning of
the content.
1.3 Inaccessibility to Basque Contents
We have started to develop a translator from Basque
to LSE because of two main reasons:
On the one hand, in the Basque Country there are
two main spoken languages: Spanish and Basque.
Not all deaf people have the ability to read and for
those who are able, learning two languages is very
challenging. Hence, those who learn a spoken
language tend to choose Spanish because of its
majority-language status (Arana et al., 2007). Due to
the hegemony of the Spanish language, all the
audiovisual content and information in Basque is out
of reach for the deaf community. Therefore, there is
a need to provide tools to make all these Basque
contents accessible to the deaf community.
On the other hand, while research on translation
from Spanish to LSE is being already studied and
developed by other research groups, no research has
been made on Basque translation into LSE. This
project takes this challenge to make research on
Basque translation and analysis.
2 RELATED WORK
At an international level, the following research
works are the most relevant ones at trying to create a
translation platform from spoken language to Sign
Language using virtual characters:
The European project ViSiCAST (Verlinden et.
al., 2001) and its continuation eSIGN (Zwiterslood
et. al., 2004) are among the first and most significant
projects related to Sign Language translation. It
consists of a translation module and a sign
interpretation module linked to a finite sign data-
base. The system translates texts of very delimited
domains of spoken language into several Sign
Languages from the Netherlands, Germany and the
United Kingdom. To build the sign data-base, a
motion capture system was used and some captions
were edited manually when needed. The result was
good regarding the execution of the signs and the
technology of that period. However, facial and body
expression were not taken into account, which is one
of the most fundamental aspects to understand Sign
Language correctly.
Dicta-Sign is a more recent European project
(Efthimiou et. al., 2010). Its goal was to develop
technologies to allow interaction in Sign Language
in a Web 2.0 environment. The system would work
in two ways: users would sign to a webcam using a
dictation style; the computer would interpret the
signed phrases, and an animated avatar would sign
the answer back to the user. The project is being
developed for the Sign Languages used in England,
France, Germany and Greece. In addition authors
affirm that their internal representation allows them
to develop a translator among different sign
languages.
There have been other smaller projects designed
to resolve local problems for deaf people in different
countries, each project working with the Sign
Languages from: United States (Huenerfauth et. al.,
2008), Greece (Efthimiou et. al., 2004), Ireland
(Smith et. al., 2010), Germany (Kipp et. al., 2011)
Poland (Francik and Fabian, 2002) or Italy
(Lombardo et. al., 2011). All of these projects try to
solve the accessibility problems that deaf people
have to access media information, communicate
with others, etc.
Regarding LSE, there have been several attempts
to build an automatic translator. Unfortunately, none
of them seem to have a huge repercussion on the
Spanish deaf community. The Speech Technology
Group at Universidad Politécnica de Madrid have
been working on a both way translation-system from
written Spanish into LSE (San-Segundo et. al.,
ProvidingAccessibilitytoHearing-disabledbyaBasquetoSignLanguageTranslationSystem
257
2006) as well as translation from LSE into written
Spanish (Lopez et al., 2010). The system is
designed for a very delimited usage domain: the
renewal of Identity Document and Driver’s license.
In a more challenging project Baldassarri et. al.
(2009) worked on an automatic translation system
from Spanish language to LSE performed by a
virtual interpreter. The system considers the mood of
the interpreter modifying the signs depending
whether the interpreter is happy, angry, etc. A more
recent research translates speech and text from
Spanish into LSE where a set of real-time
animations representing the signs are used (Vera et.
al., 2013). The main goal is to solve some of the
problems that deaf people find in training courses.
Finally, the project textoSign (www.textosign.es) is
a more oriented product whose goal is to provide a
translation service working in real time which may
be integrated into websites, digital signage
scenarios, virtual assistants, etc.
The objective of our first prototype is to provide
a domain specific spoken language-LSE translation
platform. Once its performance is validated, it could
be easily adapted to other domains in further
developments. The prototype has been built and
tested on the weather domain. This domain was
chosen for (1) using a relatively small and
predictable vocabulary, (2) having just one speaker
and (3) showing graphic help such as weather maps
as a cue for potential mistranslation cases. The first
prototype introduces two novelties: on the one hand,
the avatar will process the hand-gesture and the
bodily expression separately according the required
emotion; on the other hand, it will translate from
Basque, a co-official language in Spain and until the
moment of writing of this article, inaccessible for the
deaf community.
3 BASQUE TO LSE PROTOTYPE
The main goal of the prototype is to make
audiovisual content accessible to deaf people in a
domain specific content, so the global architecture is
validated before further development.
The system translates the input message into
LSE with the help of a virtual avatar. Figure 1 shows
the architecture of the platform and its functionality.
The input of the platform can be of diverse origin:
television, web pages, health-care services, airports,
conferences, etc.
For the first prototype, the system only allows
text and written transcriptions of spoken language as
input; however, a speech transcription module is
Figure 1: System workflow.
needed for spoken inputs such as conferences, live
television shows or talks. Even if the development of
a Basque speech recognizer is under development,
right now we apply voice alignment technologies to
align subtitles with the audio. The system has a
modular architecture to allow integrating further
speech and voice technologies in the future. Right
now the prototype works with .srt files that are
prepared before the shows based on the autocue
scripts. Voice alignment technologies are used to
match the speech to the script and synchronize the
translation.
The current system consists of five different
modules: (1) a text-to-Sign Language translation
module, (2) a gesture capture system to create an
internal (3) sign dictionary, (4) an animation engine
and a (5) rendering module. The output of the
platform can also be very diverse depending on the
ultimate context and target audience: television,
web pages, digital signage systems, mobile devices,
etc.
3.1 Text-to-Sign Language Translation
Module
The first prototype to validate the project is domain-
specific: it automatically translates a TV weather
forecast program in Basque into LSE.
Due to the lack of annotated data and the fact
that it is impossible to gather parallel corpora in
Basque and LSE, statistic machine translation
approaches were discarded, and a Rule-Based
Machine Translation approach was chosen. The
rules were designed taking into account a corpus
from the application domain of the prototype. To do
so, a code-system has been created in order to
represent LSE signs in written strings. Each concept
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
258
that in LSE has a fixed sign has its corresponding
tag in our written representation of LSE.
3.1.1 Linguistic Analysis of the Application
Domain
Before building the rules from scratch, the language
used in the application domain was linguistically
analysed. To do so, a domain-specific corpus was
compiled ad hoc. This corpus is composed by
transcriptions of 10 weather shows from the Basque-
speaking TV channel ETB1. It contains over 8000
words (2176 distinct words and 472 lemmas),
corresponding to approximately 55 minutes of
manually transcribed contents. In order to make a
thorough linguistic analysis, the corpus was tagged
semi-automatically indicating the lemma,
morphemes and the morphosyntactic information of
each token. The morphemic information, only
reachable by a deep linguistic analysis, is of special
relevance taking into account that Basque is an
agglutinative language. This is one of the major
challenges encountered in Basque automatic
translation. The linguistic information extracted was
used to spot linguistic patterns and to build robust
translation rules.
3.1.2 Text Pre-processing
The current system runs with subtitle text as input
(.srt files). These subtitles are produced before the
show and are also used as autocue help for the TV
presenter.
00:00:30,410 --> 00:00:33,060
Atzo guk iragarritakoak baino ekaitz
gutxiago izan ziren baina gaur...
(Yesterday there were fewer storms
than we had foreseen, but...)
First, the input is pre-processed to make it
suitable for automatic translation. Subtitles as such,
without text processing, follow readability standards
that hinder automatic translation: there are no more
than 32 characters per line, not more than two lines
per screenshot, etc. Hence, these sentences need to
be reconstructed before translation. The text pre-
processing module joins and splits different
sentences taking into account the information
obtained from the tokenization and the capital
letters. Besides, the whole text is lemmatized and
tagged using an inner dictionary.
[atzo|den][gu][iragarri|ad][ekaitz][
gutxi][baina|0][gaur|den][hemen][ipar][
haizea][sartu|ad][kosta]
([yesterday|time][we][foresee|v][sto
rm][few][to_be|v][to_be|aux][but|0][tod
ay|time][north][wind][enter|v][to_be|au
x][coast])
This tagging includes grammatical remarks
(word-class, time-related word or morpheme, etc.)
and other kind of linguistic information that should
be taken into account when signing in LSE. Within
this process, sentence splitting marks are also
inserted following LSE standards. Not all Basque
sentences match with LSE sentences; subordinate
clauses, for example, must be expressed in separate
sentences in LSE. The strict word order and little
abstraction of Sign Languages makes that sentences
force these languages to build shorter and
linguistically simpler sentences. The LSE sentence
splitting process is done by rules taking into account
linguistic cues.
3.1.3 Sentence Translation
The objective of this module is to give as output a
sequence of signs that strictly follows the LSE
grammar. The translation module takes as input the
processed subtitles previously obtained. The output
is a string with tags that correspond to specific signs
in LSE:
[atzo|den][gu][ekaitz][gutxi][iragar
ri|ad][baina|0][gaur|den][hemen][ipa
r][haizea][kosta][sartu|ad]
([yesterday|time][we][storm][few][fo
resee|v][but|0][today|time][north][w
ind][coast][enter|v])
Taking into account the segmentation tags and
other morphosyntactic information, it translates the
sequence that follows Basque syntax into a string of
codes that follow the LSE linguistic rules though
pattern identification. There are three types of
translation rules according to the action they imply:
Explicitation: in LSE grammatical
information is expressed with explicit
gestures. However, in Basque it is very
common to have grammatical information in
elliptical form (e.g.: subjects, tense, objects,
etc.). This information is usually expressed by
inflected word-forms and suffixes. The
explicitation rules combine all the
grammatical information contained in the
linguistic tags and the semantic information
contained in the lemmas. Thus, the output
number of tokens, each one of them
containing one grammatical or semantic piece
of information. Example:
ProvidingAccessibilitytoHearing-disabledbyaBasquetoSignLanguageTranslationSystem
259
[[.*].*[.*|v][.*|aux_past][.*].*]
↓
[[.*].*[.*|v][.*|aux_past][yesteday|
time][.*].*]
Selection of concepts: LSE does not use many
grammatical and semantically void words that
verbal languages usually have (e.g.: articles,
grammatical words, etc.). The selection rules
select the tokens that have to be expressed in
LSE and leave out the ones that do not make
sense in this language. Example (erasing
auxiliary verbs):
[[.*].*[.*|v][.*|aux_past][yesteday|
time][.*].*]
↓
[[.*].*[.*|v][yesteday|time][.*].*].
Reorganization or syntax rules: these rules
change the order of the tokens to adjust it to
the syntactic rules of LSE. The output of this
module consists of a set of tokens that can be
translated directly into a sequence of gestures.
The reorganization rules involve splitting of
sentences according to LSE rules. Example
(SOV pattern with time cue at the beginning)
[[.*].*[.*|v][yesteday|time][.*].*]
↓
[[yesteday|time][.*].*[.*|v]]
3.2 Capture System Module
In order to translate Basque into LSE, we had to
compile a LSE data-base. To do so we have
developed a capture system combining two different
motion capture systems. It uses non-invasive
capturing methods and allows entering more sign-
entries quite easily. The system can be used by any
person but only one person can use the system in
each capture session.
3.2.1 Caption of Wrist and Hand
Movements
The majority of the full-body motion capture
systems are not accurate enough when capturing
hand movements, especially movements involving
wrist and finger movements. These movements
usually require greater precision. Given their
importance in LSE, two motion capture CyberGlove
II gloves were used, one for each hand. These gloves
allow tracking precise movements of both hand and
fingers. They connect to the server via Bluetooth,
which allows more comfortable and free movements
when signing.
3.2.2 Caption of Body Movements
As mentioned before, body movements have also a
great significance in LSE. In order to capture the
movements of the whole body, the Organic Motion
system was used. This system uses several 2D
cameras to track movements. The images are
processed to obtain control points that are
triangulated to track the position of the person that is
using the system. Thanks to this system, the person
signing does not have to wear any kind of sensors,
allowing total freedom of movement. The captured
movements result more natural and realistic.
It is important that the signs are made by Sing
Language experts for a better accuracy and
understanding. The person signing has to wear the
gloves while standing inside the Organic Motion
System at the same time.
3.2.3 Merging of Hand and Body Captions
In order to join the animations captured with both
systems it is necessary to join and process the
captions before saving them as whole signs.
Autodesk Motion Builder is used for that purpose.
This software is useful to capture 3D models in real
time and it allows creating, editing and reproducing
complex animations.
Both CyberGlove II and Organic Motion provide
plug-ins to use with Motion Builder that allows
tracking and synchronizing the movements in the
same 3D scenario in real time. For that purpose three
skeletons are needed: one for each hand and one for
the whole body. These skeletons are joined in one
unique skeleton to simulate the real movements of
the person who uses the capture system. Motion
Builder allows making it in a semi-automatic way
reducing pre-processing.
Although both CyberGlove II and Organic
Motion obtain realistic movements, it is possible that
in certain cases the capture may contain errors due to
different reasons, such as calibration or
interferences. In these cases, manual post edition of
the capture is needed. This manual edition consists
of comparing the obtained animations with the real
movements and adjusting the capture using Motion
Builder. Capture edition is always done by an expert
graphic designer.
Once the realistic animations are obtained, they
are stored in a database to feed the platform with
vocabulary in Sign Language.
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
260
3.3 Sign Dictionary
The sign or gesture dictionary contains the words
used in the code given to each concept linked to the
actual gesture that the avatar has to interpret. The
gesture dictionary is composed by a finite number of
lemmatized concepts gathered from the domain-
specific corpus. Furthermore, all synonyms are
gathered within the same entry. The sign dictionary
can contain three types of entries:
One-to-One Concepts: concepts that match a
word-token in Basque and that are expressed
in one sign in LSE. Synonyms are listed under
the same LSE sign.
Grammatical or void Words: these entries are
listed in the dictionary as evidence of
processing, but are linked to an empty
concept. They do not trigger any kind of
movement because in LSE they do not exist.
Multi-word Concepts: some concepts may
map to more than one word-token in Basque.
These concepts are registered as one entry in
the gesture dictionary and they map to just one
concept in LSE.
The current sign dictionary contains 472
lemmas. These entries have proved to be enough to
translate the domain-specific corpus used to extract
the translation rules. All these concepts or lemmas
need to be captured with Capture System Module so
they are added to the sign dictionary and can be
interpreted by the virtual interpreter
3.4 Animation Engine
As explained in the second section, there are several
research projects involving a signing virtual
character. Our Animation Engine is developed with
the aim of providing natural transitions between
signs as well as modifying the execution of the signs
depending on the emotion of the virtual interpreter.
Emotion is essential in LSE. Each sign should be
represented using not only the hands and the face,
but at least, also the upper body of the interpreter. It
is based on executing the corresponding sign and
changing the speed of the animation depending on
the emotion that the virtual interpreter has to
reproduce according to the real input at that moment.
The Animation Engine runs as follows: the
appearance of the virtual interpreter is loaded from
the Virtual Character database. While the virtual
interpreter does not receive any input it has a natural
behaviour, involving blinking, looking sideways,
changing the weight of the body between both feet,
crossing arms, etc. When the Text to Sign Language
module sends the translation to the Avatar Engine
module, it stops the natural behaviour (except
blinking) and starts the sequence of signs. If any
emotion or mood cue is registered as input, the
speed of the animation changes accordingly; for
example, it slows down if sad, speeds up if angry.
Additionally, the virtual interpreter’s expression is
also modified using morphing techniques. For the
first prototype, the emotion is indicated manually
choosing one of the six universal emotions defined
by Ekman (1993), since the automatic emotion
recognition module is at a very early stage of
development.
The Animation Engine module is developed
using Open Scene Graph. It applies any sign
animations stored in Sign Language database
captured with the Capture System Module
previously. In order to concatenate several
animations and to obtain realistic movements, a
short transition between the original signs is
introduced. The implemented algorithm takes into
account the final position of a sign and the initial
position of the next sign; it is implemented using the
technique explained by Dam et., al. (1998).
Thus, the final result is the virtual interpreter
signing with very realistic movements. The
execution of the signs, as well as facial expression,
is modified depending on the emotion. Figure 2
shows the same sign executed with different
emotions.
Figure 2: The same sign with different emotions. From left
to right: neutral, angry and sad.
3.5 Rendering
The objective of this module is to visualize the
virtual interpreter synchronized with other possible
multimedia contents. For the current prototype, this
module inserts the avatar in the broadcasted TV
show. For synchronizing the virtual interpreter with
the visual content the system takes into account the
time stamps in the .srt subtitles.
ProvidingAccessibilitytoHearing-disabledbyaBasquetoSignLanguageTranslationSystem
261
4 EVALUATION
The evaluation of the system was carried out at two
levels: the translation module and animation module
were evaluated separately by LSE experts. This
allows spotting more easily the source of the
mistakes. The feedback of the evaluation is being
used as material for improvement.
4.1 Translation Evaluation
The text-to-LSE module was evaluated by one real
interpreter. In order to do the evaluation we used the
subtitles of four different TV weather program as
input of the module and we obtained the
corresponding string of signs (translated into
Spanish in request of the evaluator). Then the result
was sent to the real LSE interpreter. The sentences
to be evaluated showed the original sentence and its
“translation” to LSE with written codes in brackets
referring to one sign each.
Kaixo, arrasti on guztioi
ikusentzuleok.
[Hola][tardes][buenas][todos]
[Hello][afternoon][good][everybody]
PThe real interpreter sent back the result correcting
the sentences that contained mistakes.
Kaixo, arrasti on guztioi
ikusentzuleok.
[Hola][tardes][buenas][todos]
HOLA BUENAS TARDES TODOS
[Hello][good][afternoon][everybody]
There were translated 368 sentences with the text
to LSE module. The real interpreter returned 34 of
these sentences corrected. Hence, according to this
evaluation, the 90.7% of the translated sentences
follow strictly the LSE rules. Most of the mistakes
deal with the position of numbers or adjectives;
hence, according to the evaluator, the general
meaning of the sentences was well transmitted
despite these mistakes.
4.2 Animation Engine Evaluation
For the evaluation of the LSE animation engine a
hearing disabled person, an LSE interpreter and a
Basque/LSE bilingual person collaborated with us.
After watching the avatar execute some sentences,
they made several comments that are summarized as
follows:
Movements and their transition should come up
smoothly, especially when spelling.
The emotions could be recognized, but the
expression of the virtual interpreter ought to be more
exaggerated, using eyebrow, gaze and shoulder
movements. The waiting state of the avatar comes
up very natural, but should move less not to attract
unnecessary attention.
Some hand-modelled signs were not natural. It
was agreed that the movement caption system was
preferable.
The avatar should always wear dark clothes in
order to give more contrast between hands and
clothes.
In general, the evaluation of experts in LSE was
positive; they appreciated the efforts made to make
media services and other kind of information more
accessible. They gave a very positive feedback to the
first prototype. We used their assessment to improve
the animation engine in a second round. The
animation engine and spelling were improved, in
order to result more natural; the avatar’s clothes
have been changed to a darker colour; and both
expressions and emotions have also been
exaggerated.
In addition we are adding rules to interpret
questions and exclamations showing them on the
character’s face.
5 CONCLUSIONS
This prototype system is developed to translate
Basque text into LSE. This first version it is domain-
specific and its dictionary and grammar are limited.
In spite of these limitations, the evaluators high
lightened the impact this platform may have for the
deaf community: it makes Basque contents
accessible for the first time. Beside its social impact,
this project has proven the usability of Rule-Based
translation approaches in language combinations
where no parallel corpora is available. The Capture
System used has also been another of the keys for
the success of the project. Being a non-invasive
system with Bluetooth technology has allowed
recording more natural movements. The work done
on the animation engine has also given its results on
the smooth concatenation of signs and the execution
of the virtual interpreter’s expressions. The
expression of the avatar is another of the aspects that
was positively remarked by the experts, since the
system modifies it depending on the mood or
emotion that has to be expressed.
As future work, apart from integrating
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
262
technology to adapt other kind of inputs (e.g. audio
input), we are planning to add another motion
capture system to track the face of the person
signing. This will allow the signs to be even more
realistic, without the need of editing face
expressions manually. Furthermore, we are planning
to extend the Text to Sign Language module to other
domains. The system could also integrate other
languages such as Catalan and Catalan Sign
Language in further developments. Finally, we plan
to add an emotion recognition module in order to
recognize emotion from the voice of the speakers
and automatically modify the virtual interpreter
accordingly. Thanks to the modular architecture of
the project, it is relatively easy to integrate external
modules and/or reuse some modules in other
projects, multiplying the usability and impact of the
work done.
All these steps are planned to be taken under
LSE experts’ supervision and constant feedback.
ACKNOWLEDGEMENTS
This work have been evaluated by AransGi
(http://www.aransgi.org)
REFERENCES
Arana, E., Azpillaga, P., Narbaiza, B., 2007. Linguistic
Normalisation and Local Television in the Basque
Country. In Minority Laguage Media: Concepts,
Critiques and Case Studies. Multilingual Matters.
Noth York, 2007.
Baldassarri, S., Cerezo, E., Royo-Santas, F., 2009
Automatic Translation System to Spanish Sign
Language with a Virtual Interpreter. In Proceedings of
the 12th IFIP TC 13 International Conference on
Human-Computer Interaction: Part I, Uppsala,
Sweden: Springer-Verlag, 2009, pp. 196-199.
CESyA. 2010. Ley General Audiovisual. Available at:
http://www.cesya.es/es/normativa/legislacion/ley_gene
ral_audiovisual (accessed September 11, 2013).
Dam, E., Koch, M., Lillholm, M., 1998. Quaternions,
Interpolation and Animation, Technical Report DIKU-
TR-98/5, Department of Computer Science, Univ.
Copenhagen, Denmark.
Ekman, P., 1993. Facial expression and emotion. In
American Psychologist, Vol. 48(4) pp 384–392.
Efthimiou, E., Sapountzaki, G., Karpouzis, K. Fotinea, S-
E., 2004. Developing an e-Learning platform for the
Greek Sign Language. In Lecture Notes in Computer
Science (LNCS) , in K. Miesenberger, J. Klaus, W.
Zagler, D. Burger, Springer (eds), Vol. 3118, 2004,
1107-1113.
Efthimiou E., Fotinea S., Hanke T., Glauert J., Bowden R.,
Braffort A., Collet C., Maragos P., Goudenove F.
2010. DICTA-SIGN: Sign Language Recognition,
Generation and Modelling with application in Deaf
Communication. In 4th Workshop on the
Representation and Processing of Sign Languages:
Corpora and Sign Language Technologies (CSLT
2010), Valletta, Malta, May 2010. pp 80-84.
Francik J. and Fabian, P., 2002. Animating Sign Language
in the Real Time. In 20th IASTED International Multi-
Conference on Applied Informatics AI 2002,
Innsbruck, Austria, pp. 276-281.
Huenerfauth, M., Zhou, L, Gu, E., Allbeck, J., 2008.
Evaluation of American Sign Language generation by
native ASL signers. In ACM Trans. Access. Comput.
1, 1, Article 3.
Kipp, M., Haleoir, A. and Nguyen, Q., 2011. Sign
Language Avatars: Animation and Comprehensibility.
In Proceedings of 11th International Conference on
Intelligent Virtual Agents (IVA). Springer.
Lombardo, V., Battaglino, C., Damiano, R. and Nunnari,
F., 2011. An avatar–based interface for the Italian Sign
Language. In International Conference on Complex,
Intelligent, and Software Intensive Systems.
Lopez, V., San-Segundo, R., Martin, R., Lucas, J.M.,
Echeverry, J.D., 2010 Spanish generation from
Spanish Sign Language using a phrase-based
translation system. In VI Jornadas en Tecnología del
Habla and II Iberian SLTech Workshop, Vigo,
España.
San-Segundo, R., Barra, R., D’Haro, L.F., Montero, J.M.,
Córdoba, R., Ferreiros. J., 2006. A Spanish Speech to
Sign Language Translation System. In Proceedings of
Interspeech 2006 Pittsburgh, PA.
Smith, R., Morrissey, S., and Somers, H., 2010. HCI for
the Deaf community: Developing human-like avatars
for sign language synthesis. In Irish Human Computer
Interaction Conference’10, September 2–3, Dublin,
Ireland.
Vera, L., Coma, I., Campos, J., Martínez, B., Fernández,
M. Virtual Avatars Signing in Real Time for Deaf
Students. In GRAPP 2013 – The 8
th
International
Conference on Computer Graphics Theory and
Applications.
Verlinden, M., Tijsseling, C., Frowein, H., 2001. A
Signing Avatar on the WWW. In Proceedings of the
18th International Symposium on Human Factors in
Telecommunication, Bergen, Norway. K. Nordby,
(ed.).
Zwiterslood, I., Verlinden, M., Ros, J., van der Schoot. S.,
2004. Synthetic signing for the deaf: Esign. In
Proceedings of the Conference and Workshop on
Assistive Technologies for Vision and Hearing
Impairment, Granada, Spain, July 2004.
ProvidingAccessibilitytoHearing-disabledbyaBasquetoSignLanguageTranslationSystem
263
