Multilingual Speech Processing

64 CHAPTER 3. LINGUISTIC DATA RESOURCES

In general, ELRA is not the owner of the resources, and can therefore

only set a fair price in cooperation with the owner. This cooperation in

setting the price is often based on conventional pricing methods, such as

production costs and expected revenues. The pricing must also take into

account the ELRA distribution policy, which is to always try to offer a

discounted price to its members.

Most customers join ELRA before buying the LRs (which is enforced

by the pricing policy). ELRA’s contribution to the development of research

activities has seen considerable growth, and its involvement in research and

commercial development is balanced and shows a substantial increase in

the items distributed for R&D.

3.5 Overview of Existing Language Resources

in Europe

3.5.1 European Projects

Anumber of projects in Europe have been working toward the production of

multilingual speech and language resources, many of which have become

key databases for the HLT community. Details on some of these speech

projects follow.

The SpeechDat Family:

The SpeechDat projects

11

are a set of speech data-collection efforts funded

by the European Commission with the aim of establishing databases for the

development of voice-operated teleservices and speech interfaces. Most of

the resulting databases are available via ELRA. These projects include:

•

OrienTel: This project focused on the development of language

resources for speech-based telephone applications for the Mediter-

ranean and the Middle East, roughly spanning the area between

Morocco and the Gulf States and including several variants of local

German, French, English, Arabic, Cypriote Greek, Turkish, and

Hebrew.

11

http://www.speechdat.org

3.5. EXISTING LANGUAGE RESOURCES IN EUROPE 65

•

SALA (SpeechDat Across Latin America) which can be divided into

two further projects:

•

SALA-I: Fixed telephone network in Latin America: Speech

databases were created for the purpose of training speech

recognizers that performed well in any Latin American country.

The databases covered all dialectal regions of Latin America

that were representative in terms of Spanish and Portuguese

language variants.

•

SALA-II: Mobile/cellular telephone network in LatinAmerica,

the United States, and Canada: Speech databases were cre-

ated to train speech recognition systems for various cellphone

applications in the Americas. The databases cover all dialectal

variants of English, French, Portuguese, and Spanish languages

represented in North and Latin America.

•

SpeechDat-Car: This project focused on the development of ten in-

vehicle and mobile telephone network databases, each of which

contains 600 recording sessions. The ten languages covered were

Danish, British English, Finnish, Flemish/Dutch, French, German,

Greek, Italian, Spanish, and American English.

•

SpeechDat(E): This project aimed to provide hitherto nonexistent

resources for SpeechDat scenarios in Eastern European languages,

such as Czech, Hungarian, Polish, Russian, and Slovak.

•

SpeechDat(II): Twenty-five fixed and mobile telephone network

databases and three speaker-verification databases were developed

for Danish, Dutch, English, Flemish, Finnish, French, Belgian

French, Luxemburgian French, Swiss French, German, and Lux-

emburgian German.

•

SpeechDat(M): Eight fixed telephone network databases and one

mobile telephone network database were developed for Danish,

English, French, Swiss French, German, Italian, Portugese, and

Spanish.

Further projects related to the SpeechDat family are:

•

LILA: The goal of this recently started project is to collect a large

number of spoken databases for training automatic speech recogni-

tion systems for the Asian-Pacific languages, such as those found in

66 CHAPTER 3. LINGUISTIC DATA RESOURCES

Australia, China, India (including English), Indonesia, Japan, Korea,

Malaysia, New Zealand, the Philippines, Taiwan, Thailand, Vietnam,

etc. The data will be collected via mobile phone networks.

•

SPEECON:The overall goal of SPEECON was to enable each partner

of the consortium to produce voice-driven interfaces for consumer

applications for a wide variety of languages and acoustic environ-

ments. The languages covered by the project are Cantonese (China,

Hong Kong), Czech, Danish, Dutch (Belgium and Netherlands),

Finnish, Flemish, French, German, Hebrew, Hungarian, Italian,

Japanese, Korean, Mandarin Chinese, Russian, Spanish (American),

Swedish, Swiss-German, Turkish, U.K.-English, and U.S.-English.

Other Key European and Collaborative Projects

Producing LRs

A number of projects have targeted not only the development of LRs per

se but also the development of multilingual technology for human-human

or human-machine communication. Some of these are:

•

C-STAR (Consortium for Speech TranslationAdvanced Research):

12

This is a voluntary international consortium of laboratories devoted

to the development of spoken translation systems. The consortium

became official in 1991 and since then has undergone three phases of

collaborative research and development.

•

CHIL (Computers in the Human Interaction Loop):

13

This project

focuses on multimodal human interaction supported by computer

services that are delivered to people in an implicit, indirect, and

unobtrusive way. ELDA is involved in the activities related to LRs

and evaluation.

•

FAME (Facilitating Agents in Multicultural Exchange):

14

FAME

pursued innovation in the areas of augmented reality, perception of

human activities, and multiparty conversation modeling and under-

standing. In the context of the latter, language resources have been

collected for Catalan and Spanish language technology, information

retrieval, conversational speech understanding, robust multilingual

12

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

13

http://chil.server.de/servlet/is/101

14

http://isl.ira.uka.de/fame/