SYSTEM AND METHOD FOR GENERATING CHALLENGE UTTERANCES FOR SPEAKER VERIFICATION

• US 20120130714A1
• Filed: 11/24/2010
• Published: 05/24/2012
• Est. Priority Date: 11/24/2010
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: AT&T Intellectual Property I LP (AT&T, Inc.)
• Inventors: Zeljkovic, Ilija, Mishra, Taniya, Stent, Amanda, Syrdal, Ann K., Wilpon, Jay
• Granted Patent: US 9,318,114 B2
• Time in Patent Office: Days
• US Class Current: 704/235
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 15/08 Speech classification or search
• CPC Class Codes: G10L 17/00 Speaker identification or verification techniques
• CPC Class Codes: G10L 17/04 Training, enrolment or model building
• CPC Class Codes: G10L 17/24 the user being prompted to utter a password or a predefined phrase
• CPC Class Codes: G10L 17/26 Recognition of special voice characteristics, e.g. for use in lie detectors; Recognition of animal voices
• CPC Class Codes: G10L 2015/025 Phonemes, fenemes or fenones being the recognition units

Abstract

Disclosed herein are systems, methods, and non-transitory computer-readable storage media relating to speaker verification. In one aspect, a system receives a first user identity from a second user, and, based on the identity, accesses voice characteristics. The system randomly generates a challenge sentence according to a rule and/or grammar, based on the voice characteristics, and prompts the second user to speak the challenge sentence. The system verifies that the second user is the first user if the spoken challenge sentence matches the voice characteristics. In an enrollment aspect, the system constructs an enrollment phrase that covers a minimum threshold of unique speech sounds based on speaker-distinctive phonemes, phoneme clusters, and prosody. Then user utters the enrollment phrase and extracts voice characteristics for the user from the uttered enrollment phrase. The system generates a user profile, based on the voice characteristics, for generating random challenge sentences according to a grammar.

Claims

1. A method of performing speaker verification, the method comprising:

receiving an identity of a first user from a second user;

based on the identity, accessing a user profile comprising voice characteristics;

generating a challenge sentence, based on the voice characteristics, wherein the challenge sentence is generated randomly according to at least one of a rule and a grammar;

prompting the second user to speak the challenge sentence to yield a spoken challenge sentence; and

verifying that the second user is the first user if voice characteristics of the spoken challenge sentence are a sufficient match with the voice characteristics of the user profile.

2. The method of claim 1, wherein the challenge sentence is generated to maximize speaker discriminatory ability of the spoken challenge sentence while minimizing a length of the spoken challenge sentence.

3. The method of claim 1, wherein prompting the second user to speak the challenge sentence occurs via at least one of a text display and a text-to-speech voice.

4. The method of claim 1, wherein verifying that the second user is the first user is performed as part of a multi-platform automatic speech recognition engine.

5. The method of claim 1, wherein verifying that the second user is the first user is performed by a cloud-based speech engine shared over multiple speech service applications.

6. The method of claim 1, wherein the voice characteristics are personal level voice characteristics.

7. The method of claim 1, wherein the voice characteristics are general level voice characteristics.

8. The method of claim 1, wherein the challenge sentence is generated according to the grammar to sound semantically correct without conveying meaningful semantic information.

9. The method of claim, 1, wherein the challenge sentence is constrained to a maximum length based on an average user memory span.

10. The method of claim 1, wherein generating the challenge sentence further comprises:

identifying a slot in a sentence framework;

retrieving a word for the slot based on a personalized lexicon for the user; and

inserting the word into the slot as part of the challenge sentence.

11. A system for enrolling a user for speaker verification, the system comprising:

a processor;

a first module configured to control the processor to construct at least one enrollment phrase based on speaker-distinctive phonemes, phoneme clusters, and prosody, wherein the at least one enrollment phrase covers a minimum threshold of unique speech sounds;

a second module configured to control the processor to prompt the user to utter the at least one enrollment phrase to yield at least one uttered enrollment phrase;

a third module configured to control the processor to extract voice characteristics for the user from the at least one uttered enrollment phrase; and

a fourth module configured to control the processor to generate a user profile based on the voice characteristics, wherein the user profile is used to generate a random challenge sentence according to a challenge grammar.

12. The system of claim 11, wherein the at least one enrollment phrase is generated to maximize speaker discriminatory ability while minimizing length of the at least one enrollment phrase.

13. The system of claim 11, wherein the second module is further configured to prompt the user to utter the at least one enrollment phrase via at least one of a text display and a text-to-speech voice.

14. The system of claim 11, wherein the user profile is associated with a multi-platform, cloud-based speech engine.

15. The system of claim 11, wherein the at least one enrollment phrase is randomly generated and unpredictable.

16. A non-transitory computer-readable storage medium storing instructions which, when executed by a computing device, cause the computing device to perform steps comprising:

collecting user speech uttered while the user is using a speech service associated with a multi-platform automatic speech recognition engine;

identifying, from the user speech, voice characteristics for the user, the voice characteristics comprising speaker-distinctive phonemes, phoneme clusters, and prosody; and

generating a user profile based on the voice characteristics, wherein the user profile is used to generate a random challenge sentence according to a challenge grammar to maximize speaker discriminatory ability of the random challenge sentence while minimizing a length of the random challenge sentence.

17. The non-transitory computer-readable storage medium of claim 16, further comprising:

collecting additional user speech;

identifying additional voice characteristics from the additional user speech; and

updating the user profile based on the additional voice characteristics.

18. The non-transitory computer-readable storage medium of claim 16, wherein the voice characteristics are associated with different intonation patterns.

19. The non-transitory computer-readable storage medium of claim 16, further comprising:

identifying speech of interest to maximizing speaker discriminatory ability of the random challenge sentence;

identifying voice characteristics of interest from the speech of interest; and

flagging the voice characteristics of interest as important in the user profile.

20. The non-transitory computer-readable storage device of claim 16, wherein the random challenge sentence is generated according to one of a plurality of sentence frames.