Experimental evaluation of text-dependent speaker verification on laboratory and field test databases in the M2VTS project

This paper describes experiments of a text-independent speaker verification method that has been evaluated on two laboratory databases (M2VTS and XM2VTS) and on one field test database (LoCoMic). This work has been performed within the European ACTS-M2VTS project (Multi-Modal Verification for Teleservices and Security Applications) which is concerned with person authentication using multiple modalities. The system achieved good performance on M2VTS and XM2VTS databases whereas the performance decreased on the LoCoMic database which is more representative of real conditions. In fact, we have shown that in text independent mode, the performance differs significantly according to the test item considered (digits, name, and sentence). This shows that, even in text in dependent mode, the access control system should propose a structure for the uttered sentence.