A speaker verification system for access-control

This paper reports on the development of an automatic speaker verification system (ASV-system) integrated into an access-control environment. The development was characterized by a strong feedback loop from extensive out-of-the-lab tests into the R&D program, leading to an ASV-system with a type-I/type-II error rate crossover of < 1%, measured in the field. The ASV system is based on a word recognition approach originally devised by Thomas B. Martin [1]. It is word-dependent and requires a variable number of discrete utterances to be spoken during an access control transaction. In addition to speech algorithm quality, the overall system performance was found to be strongly determined by user interface issues, in particular the microphone type and the enrollment strategy.