Voice-based multimodal speaker authentication using adaptive training and applications thereof

1. A voice-based multimodal speaker authentication method for verifying the identity of a user of telecommunications services comprising the steps of:

adaptively training a model of an authorized user of the telecommunications services; said adaptively training comprising the steps of:

extracting a phoneme independent background model from background data resulting from a diverse set of speakers;

generating a phoneme-specific background model from the diverse set of speakers;

generating a phoneme-specific speaker model for the authorized user;

comparing the phoneme-specific background model and the phoneme-specific speaker model; and

determining from the comparison relative discriminative abilities of the two models;

establishing a telecommunications service between one or more users of the telecommunications service wherein the one or more users employ a voice-based communications method;

determining after establishment of the telecommunications service between the one or more users whether the one or more users is the authorized user of the telecommunications services by authenticating the voice of the user; and

disestablishing the established telecommunications service if the voice of the user is not the authorized user;

identifying a number of distinguishing phonemes for the two models;

generating text rich in the distinguishing phonemes; and

prompting the authorized user with the generated text to further refine the models.

2. The authentication method of claim 1 wherein said discriminative abilities of the two models is determined by measuring an Earth Movers Distance.

3. The authentication method of claim 2 wherein said determining step further comprising the steps of:

identifying one or more phrases containing distinguishing phonemes for the authorized user;

using the identified phrases to determine whether one or more of the users is the authorized user.

4. The authentication method of claim 3 further comprising the step of prompting one or more of the users to speak the identified phrases.

5. A voice-based multimodal speaker authentication method for verifying the identity of an individual comprising the steps of:

adaptively training a model of the individual wherein said adaptive training comprises the steps of:

extracting a phoneme independent background model from background data resulting from a diverse set of speakers;

generating a phoneme-specific background model from the diverse set of speakers; and

generating a phoneme-specific speaker model for the individual;

comparing the phoneme-specific background model and the phoneme-specific speaker model; and

determining from the comparison relative discriminative abilities of the two models; and

generating text rich in the distinguishing phonemes; and

identifying a number of distinguishing phonemes for the two models

prompting the individual with the generated text to further refine the models; and

using the models to determine whether a voice is that of the individual.

6. The authentication method of claim 5 wherein said discriminative abilities of the two models is determined by measuring an Earth Movers Distance.

7. The authentication method of claim 6 wherein said determining step further comprising the steps of:

identifying one or more phrases containing distinguishing phonemes for the individual;

using the identified phrases to determine whether the voice is that of the individual.

8. The authentication method of claim 7 further comprising the step of prompting one or more individuals to speak the identified phrases.