Speaker verification with elicited speaking styles in the VeriVox project

Some experiments have been carried out to study and compensate for within-s peaker variations in speaker verification. To induce speaker variation, a s peaking behaviour elicitation software package has been developed. A 50-spe aker database with voluntary and involuntary speech variation has been reco rded using this software. The database has been used for acoustic analysis as well as for automatic speaker verification (ASV) tests. The voluntary sp eech variations are used to form an enrolment set for the ASV system. This set is called structured training and is compared to neutral training where only normal speech is used. Both sets contain the same number of utterance s. It is found that the ASV system improves its performance when testing on a mixed speaking style test without decreasing the performance of the test s with normal speech. (C) 2000 Elsevier Science B.V. All rights reserved.