Speaker verification with low storage requirements

As speaker verification continues to gain acceptance for security applications, the day is drawing near when millions of users worldwide will be using some form of speaker verification system to gain access to secure information. These accesses will most probably be via the telephone network. Within this scenario a huge data flow across long distances is expected, in order to have the speaker's templates available for verification. This paper presents initial work on reducing the data requirements for text-dependent speaker verification. It is shown how data storage and accuracy can be traded off in such systems.