Continuous Speech Recognition with Speaker Verification for Secure

This work will demonstrate a novel signal processing paradigm for voice authentication. In addition, a statistical analysis technique for the evaluation and improvement of any authentication method will be developed, along with a preliminary database to support the evaluation. This includes a method for constructing a reference identification database, and a method for continuously scanning a speaker's voice to authenticate their identity. Software will extract a predetermined set of sounds which uniquely specify certain physiological characteristics of the speaker. These sounds will be processed to obtain a set of discriminating parameters unique to a specific individual's voice, yet flexible enough to allow changes which occur during normal speech. The authentication decision will be made by a neural network using the discriminating parameters. The network will take advantage of the flexibility and reliability of the proposed techniques. The training and testing of the network be achieved using the aforementioned database. ANTICIPATED BENEFITS: The ability to reliably verify the identity of an individual based on a small, quickly computable set of parameters has numberous applicatons in the military and civilian sectors. These include enhanced communication security; operator ID verification for voice controlled equipment; and customer ID verification for bank and credit cards.