Fourth European Conference on Speech Communication and Technology

Madrid, Spain
September 18-21, 1995

Speaker classification by Neural Network for Short Utteranses Using Phoneme Groups in Farsi

J. Sheikhzadegan (1), M. Tebiani (2), M. Lotfizad (1), M. R. Roohani (3)

(1) Department of Tech. & Eng., Tarbiat Modarres University, Tehran Tehran, Iran
(2) Department of Elec. Eng., Sharif University of Technology, Tehran Tehran, Iran
(3) Department of Elec. Eng., Amir Kabir University of Technology, Tehran Tehran, Iran

In this work we propose a method for automatic text-independent speaker classification for short utterances in Farsi (persian). An efficient speaker classifier was designed and implemented by neural networks (MLPs) and just made use of only the effective phonemes divided into different groups , i.e. long vowels short vowels , nasals and some of the fricatives. This classifier was evaluated with the FARSDAT speech database, from 40 native speakers with 26 males and 14 females. The results showed that the accuracy of the speaker classification with just the subclassifier related to the long vowels was over 82%, with the subclassifiers related to all vowels was over 92%, and with taking account of all of the whole classifier and considerable amount of testing process data was about 100%. Keywords: speaker classification, neural network, effective phonemes, phoneme groups, Farsi.

Full Paper

Bibliographic reference.  Sheikhzadegan, J. / Tebiani, M. / Lotfizad, M. / Roohani, M. R. (1995): "Speaker classification by neural network for short utteranses using phoneme groups in Farsi", In EUROSPEECH-1995, 375-378.