Speaker recognition

• US 9,626,971 B2
• Filed: 09/20/2013
• Issued: 04/18/2017
• Est. Priority Date: 09/28/2012
• Status: Active Grant

• Current Assignee: Cirrus Logic Incorporated
• Original Assignee: Cirrus Logic International Semiconductor Ltd. (Cirrus Logic Incorporated)
• Inventors: Artiaga, Antonio Miguel, Rodriguez, Luis Buera, Aviles-Casco, Carlos Vaquero, Gomar, Marta Garcia
• Primary Examiner(s): GUERRA-ERAZO, EDGAR X
• Application Number:
  US14/119,156
• Publication Number:
  US 20160225374A1
• Time in Patent Office: 1,306 Days
• Field of Search: None
• CPC Class Codes: G10L 17/04 Training, enrolment or model building
• CPC Class Codes: G10L 17/10 Multimodal systems, i.e. based on the integration of multiple recognition engines or fusion of expert systems
• CPC Class Codes: G10L 17/14 Use of phonemic categorisation or speech recognition prior to speaker recognition or verification
• CPC Class Codes: G10L 17/16 Hidden Markov models [HMM]
• CPC Class Codes: G10L 17/24 the user being prompted to utter a password or a predefined phrase
• CPC Class Codes: G10L 2015/025 Phonemes, fenemes or fenones being the recognition units

Abstract

Method for text-dependent Speaker Recognition using a speaker adapted Universal Background Model, wherein the speaker adapted Universal Background Model is a speaker adapted Hidden Markov Model comprising channel correction.

Claims

1. A method for text-dependent Speaker Recognition using a speaker model obtained by adaptation of a Universal Background Model,

wherein the speaker model is a speaker adapted Hidden Markov Model,

wherein the speaker model uses Bayesian inference to link observed parameters and hidden parameters,

wherein the observed parameters are the feature vectors x_nmt of utterance m of speaker n and time index t, and

wherein the hidden parameters are at least one of a group of:

the speaker factor y_n for each speaker n,

the channel factors u_nm of the utterance m of speaker n,

the active state s_nmt generating the feature vector x_nmt, and

the active component z_nmt generating the feature vector x_nmt.

2. The method for text-dependent Speaker Recognition according to claim 1, wherein the Universal Background Model is unsupervised adapted based on enrolment utterances of the speaker.

3. The method for text-dependent Speaker Recognition according to claim 1, wherein only mean vectors and transition probabilities are adapted in the speaker model or wherein all parameters are adapted in the speaker model.

4. The method for text-dependent Speaker Recognition according to claim 1, wherein the Universal Background Model of the text-dependent system is trained in an unsupervised training before it is adapted.

5. The method for text-dependent Speaker Recognition according to claim 1, wherein utterances of a plurality of speakers, which may speak more than 5 different languages are used for an unsupervised training of the Universal Background Model of the text dependent system.

6. The method for text-dependent Speaker Recognition according to claim 1, wherein the topology of the Universal Background Model of the text-dependent system is selected to comprise a transition possibility from each possible state to itself and each possible other state.

7. The method for text-dependent Speaker Recognition according to claim 1, wherein the number of states is set to a number estimated by an analysis of the spectral properties of a signal.

8. The method for text-dependent Speaker Recognition according to claim 1, further comprising adapting one or more parameters to a lexical content.

9. The method for text-dependent Speaker Recognition according to claim 1, wherein the eigenvoices matrix and eigenchannel matrix are trained from the generic Universal Background Model in a development session.

10. The method for text-dependent Speaker Recognition according to claim 1, further comprising the step of verifying in an unsupervised way whether a test signal was spoken by a target person.

11. The method for text-dependent Speaker Recognition according to claim 1, wherein the speaker adapted model is used only to determine the most likely path, but not to compute the statistics, which are useable to extract the log likelihood ratios, wherein the channel may be compensated.

12. The method for text-dependent Speaker Recognition according to claim 1, wherein verifying whether the test signal was spoken by the targeted person comprises calculating the difference between

the two terms of the log likelihood of the testing audio and the speaker model and the log product of the transition probabilities of the most likely path obtained with the speaker model

and

the log likelihood of the testing audio and the generic Universal Background Model and the log product of the transition probabilities of the most likely path obtained with the generic Universal Background Model.

13. The method for text-dependent Speaker Recognition according to claim 1, wherein the method further comprises identifying a target person by identifying the speaker adapted model with the highest likelihood score.

14. The method for text-dependent Speaker Recognition according to claim 1, wherein the Universal Background Model is a Hidden Markov Model.

15. The method for text-dependent Speaker Recognition according to claim 1, wherein the mean vectors and the transition probabilities of the Universal Background Model are adapted for the speaker model using a Maximum A Posteriori adaptation.

16. The method for text-dependent Speaker Recognition according to claim 1, wherein the channel factors are compensated in the speaker adapted model.

17. The method for text-dependent Speaker Recognition according to claim 1, wherein the following variables are used in the complete model:

a sequence of speaker factors Y

a sequence of channel factors U

a sequence of the feature vectors X

a sequence of Hidden Markov Model states S

a sequence of Gaussian components Z.

18. The method for text-dependent Speaker Recognition according to claim 1, wherein the dependencies of the variables are described by a Bayesian network.

19. The method for text-dependent Speaker Recognition according to claim 1, wherein an iterative Expectation Maximization algorithm is applied for the training of the Universal Background Model given the development data.

20. The method for text-dependent Speaker Recognition according to claim 19, wherein in the iterative algorithm in some of the iterations an additional step is introduced for maintaining boundary conditions or a step is replaced by a step for maintaining boundary conditions.

21. The method for text-dependent Speaker Recognition according to claim 1, wherein a speaker dependent Hidden Markov Model is created by adapting the mean vectors and the eigenvoice matrix of the Universal Background Model according to the enrollment data.

22. The method for text-dependent Speaker Recognition according to claim 1, wherein for the training of the Universal Background Model the model is initialized with values found by training a full covariance Universal Background Model.

23. The method for text-dependent Speaker Recognition according to claim 1, wherein the method is used for speaker verification.

25. The method for text-dependent Speaker Recognition according to claim 24, wherein text-dependent speaker recognition according to claim 1 is used.

29. A non-transitory computer readable medium comprising computer readable instructions for executing a method according to claim 1.

30. A system comprising means for executing a method according to claim 1.

31. The method of claim 1, further comprising channel correction.

32. The method of claim 2, wherein statistics are extracted using enrolment feature vectors, the Universal Background Model and a suitable algorithm.

33. The method of claim 32, wherein the suitable algorithm includes one of a forward-backward algorithm or a Viterbi algorithm.

34. The method of claim 8, wherein the one or more parameters includes at least one of the number of states or the Universal Background Model of the text dependent system.

35. The method of claim 8, wherein the lexical content is a passphrase.

37. The method of claim 1, wherein x_nmt depends on y_n, u_nm, s_nmt and z_nmt, z_nmt depends on s_nmt and s_nmt depends on s_nmt−

1.

38. The method for text-dependent Speaker Recognition according to claim 1, wherein the observation probability distributions of the Universal Background Model of the text-dependent system are Gaussian Mixture Models.

39. The method for text-dependent Speaker Recognition according to claim 38, wherein the observation probability distributions of the Universal Background Model of the text-dependent system are Gaussian Mixture Models with diagonal covariance matrices and four components.

40. The method for text-dependent Speaker Recognition according to claim 1, wherein the number of states of the Universal Background Model of the text-dependent system is fixed.

41. The method for text-dependent Speaker Recognition according to claim 20, wherein the additional step is a step of divergence re-estimation.

42. The method for text-dependent Speaker Recognition according to claim 22, wherein, after the model is initialized, the eigenvoice matrix is initialized and the eigenchannel matrix is assumed as 0.

43. The method for text-dependent Speaker Recognition according to claim 42, wherein, after the model is initialized, the eigenvoice matrix is initialized and the eigenchannel matrix is assumed as 0, then a suitable algorithm is applied neglecting any channel factors, wherein the values thus found are used to initialize the eigenchannel matrix.

44. The method for text-dependent Speaker Recognition according to claim 1, wherein the method is used for speaker identification.

24. A method for text-dependent Speaker Recognition using a text-dependent and a text-independent system, wherein a model for the text-dependent system is adapted in an unsupervised way, and wherein, in addition, a model for the text-independent system for the speaker and the phrase is built,

wherein the model uses Bayesian inference to link observed parameters and hidden parameters,

wherein the observed parameters are the feature vectors x_nmt of utterance m of speaker n and time index t, and

wherein the hidden parameters are at least one of a group of:

the speaker factor y_n for each speaker n,

the channel factors u_nm of the utterance m of speaker n,

the active state s_nmt generating the feature vector x_nmt, and

the active component z_nmt generating the feature vector x_nmt.

26. The method for text-dependent Speaker Recognition according to claim 24, further comprising the step of verifying in an unsupervised way whether a test signal was spoken by the target person.

27. The method for text-dependent Speaker Recognition according to claim 24, wherein the method further comprises a step of identifying a target person by identifying the speaker adapted model with the highest likelihood score.

28. The method for text-dependent Speaker Recognition according to claim 24, wherein the scalar weights for the linear combination have been trained in a development session.

36. The method of claim 26, wherein verifying whether the test signal was spoken by the target person comprises comparing a linear combination of the log likelihood ratios of the text-dependent system and the log likelihood ratio of the text-independent system with a threshold.

45. The method for text-dependent Speaker Recognition using a text-dependent and a text-independent system, according to claim 24, wherein the model for the text-dependent system is a Hidden Markov Model.

46. The method for text-dependent Speaker Recognition using a text-dependent and a text-independent system, according to claim 24, wherein the model for the text-dependent system is adapted by MAP.

47. The method for text-dependent Speaker Recognition using a text-dependent and a text-independent system, according to claim 24, wherein the model for the text-dependent system is adapted by a Bayesian algorithm.