Fourth European Conference on Speech Communication and Technology

Madrid, Spain
September 18-21, 1995

Discrimination of Voices of Twins and Siblings for Speaker Verification

M. Mehdi Homayounpour (1,2), Gerard Chollet (3,4)

(1) CNRS/URA 1027, Paris, France
(2) Amirkabir University of technology, Tehran, Iran
(3) IDIAP, C.P. 592, Martigny, Switzerland
(4) CNRS URA-820, ENST, Paris, France

Some experiments were conducted to study the possibility of discriminating voices of identical twins in text-independent speaker verification. Three approaches to speaker verification were examined: i) by human listeners, ii) by comparison of long-term spectra, and iii) by automatic methods [1]. Speaker verification experiments were carried out using LVQ3 and a Second Order Statistical Measure (SOSM). The results showed that listeners familiar with the voices of identical twins outperform our automatic speaker verification systems in discriminating between them. On the other hand, listeners unfamiliar with the twins perform worse than our automatic systems. This result may be explained by the fact that twins' relatives and their friends have received much more speech material for training than our automatic systems. A LVQ3-based system provides a more robust capacity than SOSM for discriminating two speakers wih similar voices such as twins.

Full Paper

Bibliographic reference.  Homayounpour, M. Mehdi / Chollet, Gerard (1995): "Discrimination of voices of twins and siblings for speaker verification", In EUROSPEECH-1995, 345-348.