method for talarverifiering / identification via modelling of typical and non typical characteristics.

摘要

The invention relates to a method at a speaker verification system/speaker identification system which makes possible for the system operator to find out the identity of a customer by means of analysis of a recording of the customer's speech. In speaker verification systems/speaker identification systems the amount of voice data which are collected from the customer is a decisive limit for the usefulness of the system. The more parameters a model comprises, the better it can be adapted to given training data at the same time as there is needed more and more training data to in a reliable way be able to estimate all existing parameters. The invention utilizes pre-trained reference models in a speaker model so that one can benefit from collected data in addition to the information which the customer himself/herself speaks in his/her registration call. The central concept of the invention is to organize said reference models in a set of pro-models and anti-models, where the pro-models model a quality which the customer has, and the anti-models model a quality which the customer does not have.