Sixth European Conference on Speech Communication and Technology

Budapest, Hungary
September 5-9, 1999

GeNeSys: A Neural Network Model for Speaker Identification

B. Ruiz-Mezcua (1), R. Rodríguez-Galán (1), Luis A. Hernández-Gómez (2), Paloma Domingo-García (1), Enrique Bailly-Baillicre Gutiérrez (1)

(1) IRIS: Laboratory of Systems Integration, Computer Science Department, Universidad Carlos III, Leganés, Madrid, Spain
(2) Signals Systems and Radiocommunications Department, ETSI Telecomunicación, Universidad Politecnica de Madrid, Spain

Mathematical models have been extensively used to shape living organism behaviour. These models are based on the N-dimensional space classification for those in which the patterns may have been defined. GeNeSys neural network family has been postulated as a global, comprehensive solution that shapes an individual behaviour. This article describes the GeNeSys family and presents some theoretical results of the researches in speaker recognition. An identification/verification system voice based is proposed. This implementation can identify or verify a speaker from 30 speakers contained in a multisession database. In this paper, a speaker verification system is presented and the tasks related to the speaker verification through the speech are developed. This system is applied to multimedia database access, services and applications. To achieve this goal a previous learning process is necessary. After the training phase is finished, the speaker model is calculated and stored in a database. A speaker recognition task using the database M2VTS from ElRA is about 88%.

Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Ruiz-Mezcua, B. / Rodríguez-Galán, R. / Hernández-Gómez, Luis A. / Domingo-García, Paloma / Bailly-Baillicre Gutiérrez, Enrique (1999): "Genesys: a neural network model for speaker identification", In EUROSPEECH'99, 1007-1010.