Second International Conference on Spoken Language Processing (ICSLP'92)

Banff, Alberta, Canada
October 13-16, 1992

Speaker Verification Using Locations and Sizes of Multipulses On Neural Networks

Seiichiro Hangai, Shigetoshi Sugiyama, Kazuhiro Miyauchi

Department of Electrical Engineering, Science University Of Tokyo, Tokyo, Japan

Locations and sizes of multipulses, which is used for driving a LPC vocal tract filter in speech synthesis, are clustered and applied to speaker verification. In clustering multipulse information, a modified K-mean algorithm is adopted. Three types of neural networks which consist of input layer with 90 ports, hidden layer with 1, 2 and 5 units and output layer with one port is also considered in verification process. After enough training using multipulse information of 70 speakers' 5 vowels, 94% of correct acceptance is achieved by determining optimum threshold.

Full Paper

Bibliographic reference.  Hangai, Seiichiro / Sugiyama, Shigetoshi / Miyauchi, Kazuhiro (1992): "Speaker verification using locations and sizes of multipulses on neural networks", In ICSLP-1992, 1439-1442.