TEXT-DEPENDENT SPEAKER VERIFICATION

• US 20080195389A1
• Filed: 02/12/2007
• Published: 08/14/2008
• Est. Priority Date: 02/12/2007
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Zhang, Zhengyou, Subramaya, Amarnag
• Granted Patent: US 8,099,288 B2
• Time in Patent Office: Days
• US Class Current: 704/246
• CPC Class Codes: G10L 17/14 Use of phonemic categorisation or speech recognition prior to speaker recognition or verification
• CPC Class Codes: G10L 17/24 the user being prompted to utter a password or a predefined phrase

Abstract

A text-dependent speaker verification technique that uses a generic speaker-independent speech recognizer for robust speaker verification, and uses the acoustical model of a speaker-independent speech recognizer as a background model. Instead of using a likelihood ratio test (LRT) at the utterance level (e.g., the sentence level), which is typical of most speaker verification systems, the present text-dependent speaker verification technique uses weighted sum of likelihood ratios at the sub-unit level (word, tri-phone, or phone) as well as at the utterance level.

Claims

1. A computer-implemented process for verifying the identity of a speaker, comprising:

asking a speaker to repeat a pass phrase;

adapting a speaker-independent model to create a speaker-specific model to be used in verifying that the pass phrase is spoken by a specific speaker.

2. The computer-implemented process of claim 1, further comprising;

inputting a speech signal containing a pass phrase as spoken by a user who claims to be a specific speaker;

extracting features from the speech signal;

inputting the features and the speaker-independent model into a speech recognizer and outputting a first recognizer score;

inputting the features and the speaker-specific model into a speech recognizer and outputting a second recognizer score; and

using the first and second recognizer scores to determine if the pass phrase is spoken by the specific speaker.

3. The computer-implemented process of claim 2 wherein using the first and second recognizer scores to determine if the pass phrase is spoken by the specific speaker employs a threshold test.

4. The computer-implemented process of claim 2 adapting a speaker-independent model to create a speaker-specific model employs a maximum likelihood procedure.

5. The computer-implemented process of claim 2 wherein certain portions of the input speech signal containing the pass phrase are weighted while using the first and second recognizer scores to determine if the pass phrase is spoken by the specific speaker.

6. The computer-implemented process of claim 2 wherein the recognizer employs hidden Markov Models.

7. The computer-implemented process of claim 2 wherein a boosting procedure is employed to give certain classes of sub-units of speech that provide more discrimination in determining if the pass phrase is spoken by the specific speaker a larger weight in comparison to other classes.

8. The computer-implemented process of claim 7 wherein the sub-units of speech are at least one of:

words;

phonemes;

or

tri-phones.

9. The computer-implemented process of claim 7 wherein the boosting procedure comprises:

for a given number of iterations,

training a classifier using a training data set of features;

computing a weighted training error using the classifier;

computing the confidence in the classifier;

updating the weights of each feature in the training data set;

normalizing the weights of each feature in the training data set; and

outputting a final classifier which uses features that give greater weight to classes of sub-units of speech that provide greater discrimination in determining whether a pass phrase is spoken by a specific speaker or an imposter.

10. The computer-implemented process of claim 2 wherein the features are one of:

raw likelihoods resulting from each recognizer pass;

the difference between the raw likelihoods and the durations of employed speech sub-units.

11. The computer-implemented process of claim 2 wherein the pass phrase the user speaks is different from a password a user types in order to verify identity.

12. The computer-implemented process of claim 2 wherein the identity of the speaker is used to gain access to one of:

phone access for banking,

database services,

shopping services,

voice mail, or

access to secure equipment.

13. The computer-implemented process of claim 2 wherein the speaker-specific model is a discriminative model and corrects errors in the speaker-independent model.

14. The computer-implemented process of claim 2 wherein the speaker-independent model employs hidden Markov Models.

15. A computer-readable medium having computer-executable instructions for performing the process recited in claim 2.

16. A system for verifying the identity of speaker, comprising:

a general purpose computing device;

a computer program comprising program modules executable by the general purpose computing device, wherein the computing device is directed by the program modules of the computer program to,

adapt a speaker-independent model to create a speaker-specific model to be used in verifying the identity of a speaker;

input a speech signal containing a pass phrase as spoken by a user who claims to be a specific speaker;

extract features from the speech signal;

input the features and the speaker-independent model into a speech recognizer and outputting a first recognizer score;

input the features and the speaker-specific model into a speech recognizer and outputting a second recognizer score; and

use the first and second recognizer scores to determine if the pass phrase is spoken by the specific speaker.

17. The system of claim 16 further comprising a program module to identify the specific speaker.

18. The system of claim 16 wherein a pass phrase is automatically obtained using the speech recognizer prior to inputting a speech signal containing a pass phrase spoken by the user, and wherein the pass phrase automatically obtained is used to determine if the speaker speaking the pass phrase is spoken by the user.

19. The system of claim 18 wherein constrained grammar is used to determine if the pass phrase spoken by the speaker is spoken by the specific speaker.

20. A system for verifying that a speaker who claims to be a specific person is that specific person, comprising:

a general purpose computing device;

a computer program comprising program modules executable by the general purpose computing device, wherein the computing device is directed by the program modules of the computer program to,

input an audio signal of a phrase spoken by a speaker claiming to be a given person into a trained speech recognizer;

input speaker-independent models into the trained speech recognizer;

input speaker-specific models of the given person'"'"'s speech that were derived by adapting the speaker-independent models into the trained speech recognizer; and

use the trained speech recognizer and the input speaker-independent models and the speaker-specific models to determine if the input audio signal of the phrase was spoken by the given person.