Fourth European Conference on Speech Communication and Technology

Madrid, Spain
September 18-21, 1995

Adaptive Transforms for Speaker Recognition

Michael J. Carey (1), Graham D. Tattersall (2), Eluned S. Parris (1)

(1) Ensigma Ltd, Turing House, Chepstow, Gwent, UK
(2) Snape Signal Research, New House, Friston, Suffolk, UK

We present, in this paper, new algorithms designed to estimate linear and non-linear transformations of the feature vectors describing the short term spectrum of the speech signal. Two algorithms are investigated: the Class Difference Maximisation Algorithm (CDMA) and Class Target Algorithm (CTA) are designed to minimise the distance between a speaker's transformed feature vectors and maximise the distance between these feature vectors and feature vectors from other speakers. This is achieved by a modified form of iterative gradient algorithm designed to maintain orthogonality between the features in the transform space. We demonstrate that the algorithm is capable of improving the separation between classes taken from different speakers. We describe the application of the transformation to the features of a Hidden Markov Model based speaker recognition experiment using broad class models. Experimental results show that when the transform is used in the system the likelihood of utterances given the true speakers models is increased while that of utterances from other speakers decreases.

Full Paper

Bibliographic reference.  Carey, Michael J. / Tattersall, Graham D. / Parris, Eluned S. (1995): "Adaptive transforms for speaker recognition", In EUROSPEECH-1995, 321-324.