Threshold Estimation with Continuously Trained Models in Speaker Verification

A-priori speaker-dependent threshold setting has been revealed as a key issue for field applications in speaker verification (SV). Threshold estimation methods have to deal with the scarcity of training data and the difficulty of obtaining data from impostors in commercial applications. The lack of client data can be faced with the implementation of a continuous training procedure. In this context, the distribution of the speaker´s scores varies with the amount of training data, resulting in an increase of verification scores for clients and impostors when more speaker data is added to the speaker model. To preserve false acceptance rate (FAR) in an acceptable margin for the application, we explore in this paper how to introduce in the threshold estimation method the observed relation between the amount of training data and the new set score distribution. Experiments are carried out over a database collected by the authors from a field application