Impostor Modelling Techniques for Speaker Verification Based on Probabilistic Neural Networks

The impact of two different impostor modelling techniques on the performance of a Probabilistic Neural Networks (PNNs)-based text-independent speaker verification system is studied. Depending on the technique used for background codebook construction, two versions of the system are obtained: with one universal background codebook, common to all authorized speakers, and with an individual speaker-dependent background codebook for each enrolled speaker. In particular, telephone-service applications over fixed and mobile telephone links with up to 500 users and short training times are considered. Results from experiments carried out on the SpeechDat(II)-FDB5000-Greek and the SpeechDat(M)-MDB1000-English corpora, with training times as short as 30 seconds and testing trials of about three seconds of voiced speech, are reported.