Fast, language-independent method for user authentication by voice

• US 8,645,137 B2
• Filed: 06/11/2007
• Issued: 02/04/2014
• Est. Priority Date: 03/16/2000
• Status: Expired due to Fees

• Current Assignee: Apple Inc.
• Original Assignee: Apple Inc.
• Inventors: Bellegarda, Jerome R., Silverman, Kim E. A.
• Primary Examiner(s): Armstrong, Angela A
• APPLICATION: US 20070294083A1
• Time in patent office: 2430 Days
• Field of Search: 704/250, 704/273, 704/270
• US Class Current: 704/250
• CPC Class Codes: G10L 15/07 to the speaker
• CPC Class Codes: G10L 17/04 Training, enrolment or model building
• CPC Class Codes: G10L 17/08 Use of distortion metrics or a particular distance between probe pattern and reference templates
• CPC Class Codes: G10L 17/14 Use of phonemic categorisation or speech recognition prior to speaker recognition or verification
• CPC Class Codes: G10L 17/22 Interactive procedures; Man-machine interfaces

Abstract

A method and system for training a user authentication by voice signal are described. In one embodiment, a set of feature vectors are decomposed into speaker-specific recognition units. The speaker-specific recognition units are used to compute distribution values to train the voice signal. In addition, spectral feature vectors are decomposed into speaker-specific characteristic units which are compared to the speaker-specific distribution values. If the speaker-specific characteristic units are within a threshold limit of the speaker-specific distribution values, the speech signal is authenticated.

Claims

1. A method of speech-based user authentication, comprising:

at a device comprising one or more processors and memory;

receiving a spoken utterance of a speaker;

generating a phoneme-independent matrix based on the spoken utterance, wherein the phoneme-independent matrix comprises a plurality of phoneme-independent feature vectors each having been extracted from a respective frame sampled from the spoken utterance at a sampling frequency;

decomposing the phoneme-independent matrix into multiple sets of vectors including at least a first set of vectors defining at least one speaker-specific recognition unit and a second set of vectors defining at least one content reference sequence;

computing at least one speaker-specific distribution value based on at least the speaker-specific recognition unit; and

authenticating an input speech signal based on the at least one speaker-specific distribution value.

2. The method of claim 1 wherein decomposing the phoneme-independent matrix further comprises:

applying a singular value decomposition to the phoneme-independent matrix.

3. The method of claim 1 further comprising:

generating the speaker-specific recognition unit from a singular value matrix of a singular value decomposition of the phoneme-independent matrix.

4. The method of claim 1, wherein authenticating the input speech signal based on the at least one speaker-specific distribution value further comprises:

decomposing at least one phoneme-independent spectral feature vector of the input speech signal into at least one speaker-specific characteristic unit;

comparing the at least one speaker-specific characteristic unit to the at least one speaker-specific distribution value; and

authenticating the input speech signal if the at least one speaker-specific characteristic unit is within a threshold limit of the at least one speaker-specific distribution value.

5. The method of claim 4 wherein decomposing the at least one phoneme-independent spectral feature vector of an input speech signal into at least one speaker-specific characteristic unit further comprises:

applying a singular value decomposition to the at least one phoneme-independent spectral feature vector of the input speech signal.

6. The method of claim 4, wherein the at least one phoneme-independent spectral feature vector is further decomposed into at least one content input sequence and authenticating the speech signal further comprises:

authenticating the input speech signal if the at least one content input sequence is similar to the at least one content reference sequence.

7. The method of claim 6 further comprising:

determining similarity based on a distance calculated between the at least one content input sequence and the at least one content reference sequence.

8. A method of authenticating a speech signal comprising:

at a device comprising one or more processors and memory;

receiving a spoken utterance of an unauthenticated speaker;

generating a phoneme-independent matrix based on the spoken utterance, wherein the phoneme-independent matrix comprises a first plurality of phoneme-independent spectral feature vectors each having been extracted from a respective frame sampled from the spoken utterance at a sampling frequency;

decomposing the phoneme-independent matrix into a speaker-specific characteristic unit;

comparing the at least one speaker-specific characteristic unit to at least one speaker-specific distribution value, the at least one speaker-specific distribution value previously trained by a registered speaker and generated by decomposing a second plurality of phoneme-independent feature vectors into sets of vectors including at least a first set of vectors defining at least one speaker-specific recognition unit and a second set of vectors defining at least one content reference sequence; and

authenticating the spoken utterance if the at least one speaker-specific characteristic unit is within a threshold limit of the at least one speaker-specific distribution value.

9. The method of claim 8 further comprising:

generating the speaker-specific characteristic unit from a singular value matrix of a singular value decomposition of the phoneme-independent matrix.

10. The method of claim 8 further comprising:

decomposing the second plurality of phoneme-independent feature vectors into the at least one speaker-specific recognition unit; and

computing the at least one speaker-specific distribution value from the at least one speaker-specific recognition unit.

11. The method of claim 10 further comprising:

generating the at least one speaker-specific recognition unit from a singular value matrix of a singular value decomposition of the second plurality of phoneme-independent feature vectors.

12. The method of claim 10 wherein decomposing the phoneme-independent matrix further comprises:

applying a singular value decomposition to the phoneme-independent matrix.

13. The method of claim 8 wherein decomposing the phoneme-independent matrix further comprises:

applying a singular value decomposition to the phoneme-independent matrix.

14. The method of claim 8, wherein the phoneme-independent matrix is further decomposed into at least one content input sequence and wherein authenticating the spoken utterance further comprises:

authenticating the spoken utterance if the at least one content input sequence is similar to at least one content reference sequence previously trained by the registered speaker.

15. The method of claim 14 further comprising:

determining similarity based on a distance calculated between the at least one content input sequence and the at least one content reference sequence.

16. A system for speech-based user authentication, comprising:

means for receiving a spoken utterance of a speaker;

means for generating a phoneme-independent matrix based on the spoken utterance, wherein the phoneme-independent matrix comprises a plurality of phoneme-independent feature vectors each having been extracted from a respective frame sampled from the spoken utterance at a sample frequency;

means for decomposing the phoneme-independent matrix into multiple sets of vectors including at least a first set of vectors defining at least one speaker-specific recognition unit and a second set of vectors defining at least one content reference sequence;

means for computing at least one speaker-specific distribution value based on at least the speaker-specific recognition unit; and

means for authenticating an input speech signal based on the at least one speaker-specific distribution value.

17. A system for authenticating a speech signal comprising:

means for receiving a spoken utterance of a speaker;

means for generating a phoneme-independent matrix based on the spoken utterance, wherein the phoneme-independent matrix comprises a first plurality of phoneme-independent spectral feature vectors each having been extracted from a respective frame sampled from the spoken utterance at a sampling frequency;

means for decomposing the phoneme-independent matrix into a speaker-specific characteristic unit;

means for comparing the at least one speaker-specific characteristic unit to at least one speaker-specific distribution value, the at least one speaker-specific distribution value previously trained by a registered speaker and generated by decomposing a second plurality of phoneme-independent feature vectors into sets of vectors including at least a first set of vectors defining at least one speaker-specific recognition unit and a second set of vectors defining at least one content reference sequence; and

means for authenticating the spoken utterance if the at least one speaker-specific characteristic unit is within a threshold limit of the at least one speaker-specific distribution value.

18. The system of claim 17, wherein the phoneme-independent matrix is further decomposed into at least one content input sequence and wherein the means for authenticating the spoken utterance further authenticates the spoken utterance if the at least one content input sequence is similar to at least one content reference sequence previously trained by the registered speaker.

19. The system of claim 18, wherein the means for comparing further determines similarity based on a distance calculated between the at least one content input sequence and the at least one content reference sequence.

20. A non-transitory computer readable medium comprising instructions, which when executed on a processor, perform a method of speech-based user authentication, comprising:

receiving a spoken utterance of a speaker;

generating a phoneme-independent matrix based on the spoken utterance, wherein the phoneme-independent matrix comprises a plurality of phoneme-independent feature vectors each having been extracted from a respective frame sampled from the spoken utterance at a sampling frequency;

decomposing the phoneme-independent matrix into multiple sets of vectors including at least a first set of vectors defining at least one speaker-specific recognition unit and a second set of vectors defining at least one content reference sequence

computing at least one speaker-specific distribution value based on at least the speaker-specific recognition unit; and

authenticating an input speech signal based on the at least one speaker-specific distribution value.

21. A non-transitory computer readable medium comprising instructions, which when executed on a processor, perform a method for authenticating a speech signal, comprising:

receiving a spoken utterance of a speaker;

generating a phoneme-independent matrix based on the spoken utterance, wherein the phoneme-independent matrix comprises a first plurality of phoneme-independent spectral feature vectors each having been extracted from a respective frame sampled from the spoken utterance at a sampling frequency;

decomposing the phoneme-independent matrix into a speaker-specific characteristic unit;

comparing the at least one speaker-specific characteristic unit to at least one speaker-specific distribution value, the at least one speaker-specific distribution value previously trained by a registered speaker and generated by decomposing a second plurality of phoneme-independent feature vectors into sets of vectors, including at least a first set of vectors defining at least one speaker-specific recognition unit and a second set of vectors defining at least one content reference sequence; and

authenticating the spoken utterance if the at least one speaker-specific characteristic unit is within a threshold limit of the at least one speaker-specific distribution value.

22. The computer readable medium of claim 21, wherein the phoneme-independent matrix is further decomposed into at least one content input sequence and wherein authenticating the speech signal further comprises:

authenticating the speech signal if the at least one content input sequence is similar to at least one content reference sequence previously trained by the registered speaker.

23. The computer-readable medium of claim 22, wherein the method further comprises:

determining similarity based on a distance calculated between the at least one content input sequence and the at least one content reference sequence.

24. A system for speech-based user authentication, comprising:

a processor configured to

receive a spoken utterance of a speaker,

generate a phoneme-independent matrix based on the spoken utterance, wherein the phoneme-independent matrix comprises a plurality of phoneme-independent feature vectors each having been extracted from a respective frame sampled from the spoken utterance at a sampling frequency,

decompose the phoneme-independent matrix into multiple sets of vectors at least a first set of vectors defining at least one speaker-specific recognition unit and a second set of vectors defining at least one content reference sequence,

compute at least one speaker-specific distribution value based on at least the speaker-specific recognition unit; and

authenticate an input speech signal based on the at least one speaker-specific distribution value.

25. The system of claim 24 wherein the processor is further configured to decompose the phoneme-independent matrix into the at least one speaker-specific recognition unit.

26. The system of claim 25 wherein the processor is further configured to apply a singular value decomposition to the phoneme-independent matrix to generate the at least one speaker-specific recognition unit.

27. The system of claim 24 wherein the processor is further configured to generate the at least one speaker-specific recognition unit from a singular value matrix of a singular value decomposition of the phoneme-independent matrix.

28. The system of claim 24 wherein the processor is further configured to decompose at least one phoneme-independent spectral feature vector of an input speech signal into at least one speaker-specific characteristic unit, and authenticate the speech signal if the at least one speaker-specific characteristic unit is within a threshold limit of the at least one speaker-specific distribution value.

29. The system of claim 28 wherein the processor is further configured to apply a singular value decomposition to the at least one phoneme-independent spectral feature vector of the input speech signal.

30. The system of claim 28, wherein the processor is further configured to decompose the at least one phoneme-independent spectral feature vector of the input speech signal into at least one content input sequence, and to authenticate the input speech signal if the at least one content input sequence is similar to the at least one content reference sequence.

31. The system of claim 30, wherein the processor is further configured to determine similarity based on a distance calculated between the at least one content input sequence and the at least one content reference sequence.

32. A system for authenticating a speech signal comprising:

a processor to

receive a spoken utterance of an unauthenticated speaker,

generate a phoneme-independent matrix based on the spoken utterance, wherein the phoneme-independent matrix comprises a first plurality of phoneme-independent spectral feature vectors each having been extracted from a respective frame sampled from the spoken utterance at a sampling frequency,

decompose the phoneme-independent matrix into a speaker-specific characteristic unit,

compare the at least one speaker-specific characteristic unit to at least one speaker-specific distribution value, the at least one speaker-specific distribution value previously trained by a registered speaker, and

authenticate the spoken utterance if the at least one speaker-specific characteristic unit is within a threshold limit of the at least one speaker-specific distribution value, wherein the at least one speaker-specific distribution value is generated by decomposing a second plurality of phoneme-independent feature vectors into sets of vectors including a first set of vectors defining at least one speaker-specific recognition unit and a second set of vectors defining at least one content reference sequence.

33. The system of claim 32 wherein the processor is further configured to apply a singular value decomposition to the phoneme-independent matrix.

34. The system of claim 32 wherein the processor is further configured to generate the at least one speaker-specific characteristic unit from a singular value matrix of a singular value decomposition of the phoneme-independent matrix.

35. The system of claim 32 wherein the processor is further configured to decompose the second plurality of phoneme-independent feature vectors into the at least one speaker-specific recognition unit, and compute the at least one speaker-specific distribution value from the at least one speaker-specific recognition unit.

36. The system of claim 35 further comprising:

a feature extractor to extract the second plurality of phoneme-independent feature vectors into a speaker-specific feature extraction representation.

37. The system of claim 36 wherein the processor is further configured to decompose the speaker-specific feature extraction representation into the at least one speaker-specific recognition unit.

38. The system of claim 37 wherein the processor is further configured to apply a singular value decomposition to the speaker-specific extraction representation to generate the at least one speaker-specific recognition unit.

39. The system of claim 35 wherein the processor is further configured to generate the at least one speaker-specific recognition unit from a singular value matrix of a singular value decomposition of the second plurality of phoneme-independent feature vectors.

40. The system of claim 32, wherein the phoneme-independent matrix is further decomposed into at least one content input sequence and wherein the processor is further configured to authenticate the spoken utterance if the at least one content input sequence is similar to at least one content reference sequence previously trained by the registered speaker.

41. The system of claim 40, wherein the processor is further configured to determine similarity based on a distance calculated between the at least one content input sequence and the at least one content reference sequence.