5th International Conference on Spoken Language Processing
A new discriminant speaker model is introduced in this paper. The model is text dependent and relies on characterising speakers in terms of the angular distance between ``projection vectors'', which allow good discrimination between individual speakers. The projection models require only little enrollment data to be available per target speaker, but at the same time require a set of ``cohort speakers'' to be available for which a relatively large amount of training speech is available per cohort speaker. The projection model technique is evaluated on the Gandalf database and compared to conventional Gaussian Mixture Models (GMMs). It is found that the projection models require less storage per target speaker, while at the same time achieving lower error rates, particularly when applied for speaker identification and recognition under mismatched conditions.
Bibliographic reference. Olsen, Jesper Ostergaard (1998): "Speaker recognition based on discriminative projection models", In ICSLP-1998, paper 0335.