Lietuvių bendrinės šnekamosios kalbos garsų fondo kūrimo principai

This paper presents the VDU Lithuanian speech corpus. The corpus has been compiled and annotated by the Center of Computational Linguistics at the Vytautas Magnus University. The corpus aims at providing basic data for Lithuanian language technology researches meant to enable Lithuanian spoken language researchers to use advanced tools provided by today’s computer science. The VDU speech corpus contains broadband recordings of 4 speakers (2 males and 2 females), each reading the same set of nearly 7540 isolated words and the same number of word triplets. The corpus includes timealigned phone-level, phonetic unit-level and word-level transcriptions as well. The VDU Lithuanian speech corpus is universal, i.e. its vocabulary has been carefully chosen to include all distinct and independent Lithuanian sounds such as phonemes and phoneme clusters (phonetic units). There have been 275 such phonetic units defined. The paper also describes problems related to the file structure of the corpus and the SAMPAASCII coding of Lithuanian annotations. Some other questions are discussed, such as corpus documentation, validation and standardization. These questions have been addressed in Lithuania for the first time