This paper describes a spoken language identification method based on speech fundamental frequency (F0). The procedure is subdivided into three main stages: 1) F0 extraction and segmentation; 2) polygonal line approximation of F0 pattern; 3) discriminant analysis. The stage of F0 extraction uses the Average Magnitude Difference Function(AMDF) and speech energy to estimate the fundamental frequency period of voiced speech sounds. In order to find better features from F0 pattern, polygonal lines are used to approximate the F0 contour of voiced intervals. After previous two stages, the complete parameter set is available for discrimination. The principal component analysis and discriminant analysis are performed at the last stage. The system is trained and tested using a CD-ROM. The Multi-language Speech Database for Telephonometry 1994", which is produced by NTT and NATC, and the OGI Multi-language Telephone Speech Corpus. Keywords: F0 contour, Principal Component analysis, Discriminant analysis.
Bibliographic reference. Shuichi, Itahashi / Liang, Du (1995): "Language identification based on speech fundamental frequency", In EUROSPEECH-1995, 1359-1362.