SALA: SpeechDat across Latin America. Results of the First Phase

The objective of the SALA (SpeechDat across Latin America) project is to record large SpeechDat-like databases to train telephone speech recognisers for any country in Latin America. The SALA consortium is composed by several European The Netherlands). This paper gives an overview of the project, introduces the definition of the databases, shows the dialectal distribution in the countries where recordings take place and gives information about validation issues, actual status and practical experiences in recruiting and annotating such large databases in Latin America ,QWURGXFWLRQ SALA (6peechDat $cross /atin $merica) belongs to the family of SpeechDat projects 1 producing speech databases to train recognizers for telephone applications. specification of the databases is coherent with those of the SpeechDat II databases. Some adjustments and additions of the specifications have to be made reflecting the different needs for Latin America. Also the validation procedure of the databases is performed along the rules of SpeechDat II. For validation of the databases the institute SPEX – ELRA's validation center of spoken language resources – is responsible. The goal of SALA is to produce speech databases in a way that the recognizers trained with these databases show good recognition performance for all different dialectal variants of Spanish and Portuguese as spoken in Latin America. Due to the lack of precise phonetic knowledge of the dialectal regions in Latin America the members of the SALA project decided to first collect sample databases recorded from 1000 – 2000 speakers for selected Latin American countries (phase I of the project). In a second phase of the SALA project the recordings should be enlarged for specific dialectal regions. The selection of these regions should be based on recognition experiments made with the databases produced so far. The project SALA is still in phase I. The countries so far covered are: Brazil: 2000 speakers Mexico: 2000 speakers Venezuela: 1000 speakers Colombia: 1000 speakers. Some other companies are in negotiation to join the consortium providing speech databases for Argentina. Still the dialects found in Chile or Peru are not yet covered. Due to the experiences made so far it turned out that the production of speech databases in Latin America takes much longer time than expected. Main problems which have to be solved concern • insufficient developed infrastructure of the telephone network • long lasting contractual and legal issues • delay through costums regulation. Now phase I of the SALA project is …