Methods and apparatus for verifying the originator of a sequence of operations

• US 5,526,465 A
• Filed: 04/05/1993
• Issued: 06/11/1996
• Est. Priority Date: 10/03/1990
• Status: Expired due to Term

• Current Assignee: Ensigma Ltd. (Imagination Technologies Group Ltd.)
• Original Assignee: Ensigma Ltd. (Imagination Technologies Group Ltd.)
• Inventors: Carey, Michael J., Bridle, John S., Parris, Eluned S.
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Dorvil, Richemond
• Application Number:
  US08/039,054
• Time in Patent Office: 1,163 Days
• Field of Search: 395/2.59, 395/2.65, 395/2.82, 395/2.64, 395/2.65, 395/2.55, 395/2.58, 395/2.52, 395/2.54, 382/10, 382/13, 382/14, 382/15, 382/2, 382/3, 381/41, 381/42, 381/43
• US Class Current: 704/250
• CPC Class Codes: G10L 17/00 Speaker identification or verification techniques

Abstract

Speaker verification is important in such applications as financial transactions which are to be carried out automatically by telephone. False acceptances of a speaker cause serious problems but so do frequent false rejections in view of the annoyance caused. Some of the problems of speaker verification are reduced in the invention by forming Hidden Markov Models (HMMs) for each of a number of words using features of utterances of these words from a large number of speakers. These models are known as world models. In addition, for every person whose speech is to be recognized, one HMM is formed for each of the words as uttered by that person. These models are known as personal models. In verification, a person is prompted to repeat a string of isolated or connected words (15) and features from each of these words are extracted (16). Next, the probabilities that these features could have been generated by the world models for these words and by the personal model of that person are calculated, respectively (17 and 18) and these probabilities are compared (19) for each word. A decision (23) on verification is based on a poll (22) of these comparisons.

Claims

1. A method of verifying that a sequence of operations originates from a specific entity, comprising the steps of

extracting a test sequence of sets of features of the results of the operations, one set corresponding to each operation,

matching the said test sequence of sets of features against a first stored finite state machine model derived from sets of features of the results of the same sequence of operations when originated by a plurality of entities,

matching the said test sequence of features against a second stored finite state machine model derived from sets of features of the results of the same sequence of operations when originated by the specific entity, and

comparing the results of the matching steps to indicate whether the test sequence of operations originated from the specific entity.

2. A method according to claim 1 wherein the entity is a person and the sequence of operations is the utterance of sequences of sounds.

3. A method according to claim 2 wherein each feature comprises a multi-element vector in which the elements together provide a representation of the sequence of sounds in an interval, and the set of vectors represents the sequence of sounds over a succession of the said intervals.

4. A method according to claim 2 wherein the utterance is the utterance of an item chosen from an alpha-numeric character and a word.

5. A method according to claim 4 wherein

matching the said test sequence against a first stored finite state machine model comprises matching against a selected one of a plurality of first stored finite state machines models each derived from respective utterances of an item when uttered by each of a plurality of persons, and

matching the said test sequence against a second stored finite state machine model comprises matching against that one of a plurality of second finite state machine models which is derived from the same item as the said selected one model, each second model being derived from utterances of a respective item when uttered by a specific person.

6. A method according to claim 5 wherein

comparing the results of the matching step is by comparing probabilities calculated for each of a plurality of utterances from first and second said models derived from utterances of the same items, and indicating whether the plurality of utterances was from the specific person on the basis of the majority of results of the probability comparisons.

7. A method of speaker verification including

prompting a speaker to utter a series of isolated or connected items chosen from alpha-numeric characters and words,

carrying out a method according to claim 5 for each utterance made in response to prompting.

8. A method according to claim 5 wherein matching the said test sequence includes for each of the first and second finite state machine models

calculating for each state in that finite state machine model the maximum probability that a sequence of states, obtainable from that finite state machine model and ending in that state, could generate the test sequence,

determining the end of each item in the sequence of sets of features based on the said maximum probability for each state,

using the said maximum probability for each state at the end of an item as the probability of that item, and

comparing the result of the matching steps includes comparing the maximum probabilities for each item as obtained for the first and second finite state machine models.

9. A method according to claim 2 wherein the utterance is the utterance of a string of connected spoken items chosen from alphanumeric characters and words.

10. A method according to claim 9 wherein

matching the said test sequence against a first stored finite state machine model comprises matching against a model comprising a stored item model for each item in the utterance, the item models being serially linked and each derived from respective utterances of that item when uttered by each of a plurality of persons, and

matching the said test sequence against a second stored finite state machine model comprises matching against a model comprising a stored item model for each item in the utterance, the item models being serially linked and each derived from respective utterances of that item when uttered by a specific person.

11. A method according to claim 2 wherein

each matching step comprises the calculation of a probability, and

comparing the results of the matching steps is by subtracting one log probability from another, the sign of the result indicating whether the utterance was from a specific person.

12. A method according to claim 1 wherein the entity is a person and the sequence of operations is the writing of a word or letter.

13. A method according to claim 1 wherein extracting the test sequence includes cepstral processing to provide features in the form of cepstral coefficients.

14. A method according to claim 13 wherein the zero cepstral coefficient is omitted from the sets of features.

15. A method according to claim 13 wherein the first cepstral coefficient is omitted from the sets of features.

16. A method according to claim 13, wherein extracting the test sequence includes

providing a set of power values for each feature in which each power value represents the power of sound in one of a series of equal intervals over a respective frequency band,

providing cepstral processing for the power values by calculating a logarithm value for each power value and carrying out a cosine transform on the logarithm values, and

using at least some of the resulting cepstral coefficients as respective said features.

17. A method according to claim 16 including, for at least one cepstral coefficient, determining first difference related features based on values of the said one coefficient as derived from a number of the said sets of power values which occur in different said intervals, and using the first difference related features as features of the said sets of features.

18. A method according to claim 17 including deriving a number of second difference related features based on a plurality of first difference related features derived from different values of one cepstral coefficient, and using the second difference related features as features of the said sets of features.

19. A method according to claim 18, wherein the second difference related feature is derived by calculating the difference between alternate first differences in a series derived from the series of cepstral coefficient values.

20. A method according to claim 18 wherein each second difference related feature is derived by weighting first difference related features in a series derived from the series of cepstral coefficients using respective multipliers which decrease to a minimum and increase after the minimum, forming the sum of first difference related features weighted with decreasing multipliers, forming the sum of first difference related features weighted with increasing multipliers, and forming the difference between the two sums to give the second difference related features.

21. A method according to claim 17 wherein the first difference related feature is derived by calculating the difference between alternate values of a cepstral coefficient in a series of values derived from the power values in a series of the said intervals.

22. A method according to claim 17 wherein each first difference related feature is derived by weighting values of a cepstral coefficient in a series of values derived from the power values in a series of the said intervals using respective multipliers which decrease to a minimum and increase after the minimum, forming the sum of the coefficients weighted with decreasing multipliers, forming the sum of the coefficients weighted with increasing multipliers, and forming the difference between the two sums to give the first difference related feature.

23. A method according to claim 1 wherein the finite state machines are Hidden Markov Models.

24. A method according to claim 1, wherein said specific entity is a word or phrase, and said plurality of entities comprise a plurality of different words or phrases.

25. A method according to claim 1, wherein said specific entity is a language, and said plurality of entities comprise a plurality of different languages.

26. Apparatus for verifying that a sequence of operations originated from a specific entity, comprising

means for storing data specifying first and second finite state machines, the data for the first machine having been derived from sets of features of the results of a sequence of operations originated by a plurality of entities, the data for the second machine having been derived from sets of features of the results of the same sequence of operations originated by a specific entity,

means for extracting a test sequence of sets of features from the results of a sequence of operations which are alleged to have been originated by the said specific entity,

means for matching the said test sequence against the first and second said machines, respectively, and

means for comparing results from the matching means to indicate whether the test sequence was originated by the said specific entity.

27. Apparatus according to claim 26 wherein the entity is a person and the sequences of operations are the utterances of sequences of sounds.

28. Apparatus according to claim 27 wherein the finite state machines are Hidden Markov Models.

29. Apparatus according to claim 28 wherein the data specifying the Hidden Markov Models comprises a set of pairs of means (m) and variances (σ

²) for each state, the pairs of means and variances having been derived from respective distributions of values each representing a spectral component of a sound in the sequences used to derive the Hidden Markov Models.

30. Apparatus according to claim 29 wherein the means and variances are calculated from cepstral coefficients derived from the said spectral components and first and/or difference related features derived from the cepstral coefficients.

31. Apparatus according to claim 30 wherein the zero and first cepstral coefficients are omitted when deriving the means and variances.

32. Apparatus according to claim 29 wherein the means and variances are modified means and modified variances calculated using a gradient algorithm in an alpha-net comprising the two finite state machines.

33. Apparatus according to claim 32 wherein the modified mean m_j (t) and the modified standard deviation σ

_j (t) are calculated from ##EQU15## where m_j (t-1) and σ

_j (t-1) represent the previous mean and standard deviation,

ξ

is a constant, and

J is a score value derived from the likelihood output of one of the first and second finite state machines.

34. Apparatus according to claim 32 wherein the modified mean m_j (t) and the modified standard deviation σ

_j (t) are calculated from ##EQU16## where m_j (t-1) and σ

_j (t-1) represent the previous mean and standard deviation,

ξ

is a constant, and

J is a score value derived from the likelihood output of one of the first and second finite state machines.

35. Apparatus according to claim 26 wherein the entity is a person and the sequences of operations are the signing of signatures.

36. A method of verifying that operations originate from a specific entity, comprising the steps of

extracting test results of the operations,

matching the said test results against a first model derived from results of the same operations when originated by a plurality of entities,

matching the said test results against a second model derived from results of the same operations when originated by the specific entity, and

comparing the results of the matching steps to indicate whether the test results originated from the specific entity.

37. A method of identifying an utterance, which comprises:

extracting a test sequence of sets of features for said utterance;

matching said test sequence of sets of features against a first stored finite state machine model derived from sets of features corresponding to a group of utterances;

matching said test sequence of sets of features against a second stored finite state machine model derived from sets of features corresponding to said given utterance; and

comparing the results of the matching steps to indicate whether said utterance corresponds to said given utterance.

38. A method according to claim 37, wherein said utterance comprises a word or phrase.

39. A method according to claim 37, wherein said utterance comprises a sample of language.

40. Apparatus for identifying an utterance, comprising

means for storing data specifying first and second finite state machines, the data for the first machine having been derived from sets of features corresponding to a group of utterances, and the data for the second machine having been derived from sets of features corresponding to said given utterance;

means for extracting a test sequence of sets of features for said utterance;

means for matching said test sequence against the first and second finite state machines, respectively;

means for comparing results from said matching means to indicate whether said utterance corresponds to said given utterance.