Phonetic Level Annotation and Segmentation of Hungarian Speech Databases

The purpose of this paper is to give an outline of phonetic level annotation and segmentation of Hungarian speech databases at the levels of definition and speech technology. In addition to giving guidance to the definition of the content of a database, the technique of annotation and the procedure of manual segmentation, we also discuss mathematical models of computeraided semi-automatic and automatic segmentation. Finally, we are summing up our observations about the application of the procedures we gained during the processing of the MTBA Hungarian Telephone Speech Database. 1 Designing a Speech Database Statistics based speech processing, particularly automatic speech recognition, requires well-organized, large speech databases. Training a speech recognition program is based on statistical parameter estimation. Accurate parameter tuning requires training on a large number of samples. A proper training database is made up of collections of such samples accompanied with the necessary notes, labels and transcriptions. The databases should include the observations that are required by the parameter estimation and all the samples that cover the variability of speech (and noises of the environment). A speech database is a large set of sound data which can be organized by several grouping conditions. The size and internal structure of a database is usually determined by the area of use. To achieve a reliable level of accurate recognition, the material should contain every typical variation that is likely to occur during recognition. ∗Presented at the 1st Conference on Hungarian Computational Linguistics, December 10–11, 2003, Szeged. †Department of Hungarian Linguistics, University of Szeged, H-6722 Szeged, Egyetem utca 2., Hungary, e-mail: [email protected] ‡Research Group on Artificial Intelligence of the Hungarian Academy of Sciences and University of Szeged, H-6720 Szeged, Aradi vértanúk tere 1., Hungary, e-mail: [email protected], [email protected] §Department of Hungarian Linguistics, University of Szeged, H-6722 Szeged, Egyetem utca 2., Hungary, e-mail: [email protected]