Determining distortion measures in a pattern recognition process

• US 7,912,715 B2
• Filed: 03/27/2003
• Issued: 03/22/2011
• Est. Priority Date: 03/27/2002
• Status: Expired due to Fees

• Current Assignee: WSOU Investments, LLC (WSOU Holdings, LLC)
• Original Assignee: Nokia Corporation
• Inventors: Vasilache, Marcel
• Primary Examiner(s): Opsasnick, Michael N
• Application Number:
  US10/402,371
• Publication Number:
  US 20040039573A1
• Time in Patent Office: 2,917 Days
• Field of Search: 704/251, 704/254, 704/255, 704/256.6
• US Class Current: 704/251
• CPC Class Codes: G06F 2218/00 Aspects of pattern recognition specially adapted for signal processing
• CPC Class Codes: G10L 15/10 using distance or distortion measures between unknown speech and reference templates

Abstract

A method for determining a set of distortion measures in a pattern recognition process, where a sequence of feature vectors is formed from a digitized incoming signal to be recognized, said pattern recognition being based upon said set of distortion measures. The method comprises comparing (S10) a first feature vector in said sequence with a first number (M1) of templates from a set of templates representing candidate patterns, based on said comparison, selecting (S12) a second number (M2) of templates from said template set, the second number being smaller than the first number, and comparing (S14) a second feature vector only with said selected templates. The method can be implemented in a device for pattern recognition.

Claims

1. A method comprising:

comparing by a processor a first feature vector in a sequence of feature vectors formed from a digitized incoming signal to be recognized, with a first number of templates from a set of templates representing candidate patterns,

based on said comparison, selecting by a processor in response to a control signal, a second number of templates from said template set, the second number being smaller than the first number,

comparing by a processor a second feature vector only with said selected templates, and

generating by a processor a signal corresponding to a recognized pattern of said digitized incoming signal as a result of said comparing said second feature vector only with said selected templates.

2. A method according to claim 1, wherein said second number is dependent upon a distance measure between said first feature vector and said second feature vector.

3. A method according to claim 1, wherein said selected templates include the templates resulting in the lowest distortion measures when compared to said first feature vector.

4. A method according to claim 1, wherein said selected templates include a pre-determined number of templates resulting in the lowest distortion measures when compared to said first feature vector.

5. A method according to claim 1, wherein said selected templates include all templates resulting in a distortion measure below a predefined threshold value when compared to said first feature vector.

6. A method according to claim 1, wherein a number of successive feature vectors are only compared with said second number of templates in said template set.

7. A method according to claim 1, wherein, for templates not included in said selected templates, distortion measures computed with respect to a different feature vector are included in said set of distortion measures.

8. A method according to claim 1, wherein, for templates not included in said selected templates, specific components of distortion measures computed with respect to a different feature vector are used for determining said set of distortion measures.

9. A method according to claim 7, wherein said different feature vector is the feature vector most recently compared to the first number of templates from the template set.

10. A method according to claim 8, wherein said different feature vector is the feature vector most recently compared to the first number of templates from the template set.

11. A method according to claim 7, wherein said different feature vector is a feature vector compared to the first number of templates from the template set being closest to the current feature vector according to a predefined distance measure.

12. A method according to claim 8, wherein said different feature vector is a feature vector compared to the first number of templates from the template set being closest to the current feature vector according to a predefined distance measure.

13. A method according to claim 1, wherein said number of successive feature vectors is static.

14. A method according to claim 1, wherein said number of successive feature vectors is dynamic.

15. A method according to claim 1, wherein said control signal is based on a time-dependent variable belonging to the group of processor load and incoming signal properties.

16. A method according to claim 1, wherein said templates are Gaussian mixture densities of Hidden Markov Models.

17. A method according to claim 16, wherein said distortion measures are based on log-likelihoods.

18. A method according to claim 16, wherein said pattern recognition includes computing a state likelihood for a Hidden Markov Models with respect to a feature vector.

19. A method according to claim 1, wherein said signal represents speech, and said candidate patterns represent spoken utterances.

20. A computer readable storage medium stored with code which, when executed by a processor, causes an apparatus to determine a set of distortion measures in a pattern recognition process by performing:

forming a sequence of feature vectors from a digitized incoming signal to be recognized, said pattern recognition being based upon said set of distortion measures,

comparing a first feature vector with a first number of templates from a set of templates representing candidate patterns,

based on said comparison, selecting a second number of templates from said template set, the second number being smaller than the first number,

comparing a second feature vector only with said selected templates, so as to recognize a pattern of said digitized incoming signal; and

generating a signal corresponding to a recognized pattern of said digitized incoming signal as a result of said comparing said second feature vector only with said selected templates.

21. An apparatus comprising:

a distortion computation module configured;

to compare a first feature vector in a sequence of feature vectors formed from a digitized incoming signal to be recognized with a first number of templates from a set of templates representing candidate patterns,

to select, based on said comparison, a second number of templates from said template set, the second number being smaller than the first number,

to compare a second feature vector only with said selected templates so as to recognize a pattern of said digitized incoming signal; and

to generate a signal corresponding to a recognized pattern of said digitized incoming signal as a result of said comparing said second feature vector only with said selected templates.

22. The apparatus according to claim 21, wherein said selected templates include the templates resulting in the lowest distortion measures when compared to said first feature vector.

23. The apparatus according to claim 21, further wherein the distortion computation module is configured to include distortion measures computed with respect to a different feature vector, in said set of distortion measures.

24. The apparatus according to claim 21, wherein said distortion computation module configured to compare said second feature vector is configured to compare a number of successive feature vectors only with said selected templates.

25. The apparatus according to claim 21, further comprising a control module, configured to detect the processor load and to adjust the number of successive feature vectors in response to a said load.

26. A speech recognizer comprising an apparatus according to claim 21.

27. A communication device comprising a speech recognizer according to claim 26.

28. An apparatus for pattern recognition, comprising:

a feature extractor configured to form a sequence of feature vectors from a digitized incoming signal,

a pattern recognizer configured to perform a pattern recognition process based upon a set of distortion measures, and

a distortion computation module configured;

to compare a first feature vector with a first number of templates from a set of templates representing candidate patterns,

to select, based on said comparison, a second number of templates from said template set, the second number being smaller than the first number,

to compare a second feature vector only with said selected templates, so as to recognize a pattern of said digitized incoming signal; and

to generate a signal corresponding to a recognized pattern of said digitized incoming signal as a result of said comparing said second feature vector only with said selected templates.

29. The apparatus according to claim 28, implemented as an embedded system, comprising:

a front-end section for forming said sequence of feature vectors, and

a back-end section for determining said set of distortion measures.

30. An apparatus comprising:

means for comparing a first feature vector in a sequence of feature vectors formed from a digitized incoming signal to be recognized with a first number of templates from a set of templates representing candidate patterns,

means for selecting, based on said comparison, a second number of templates from said template set, the second number being smaller than the first number, and

means for comparing a second feature vector only with said selected templates, so as to recognize a pattern of said digitized incoming signal; and

means for generating a signal corresponding to a recognized pattern of said digitized incoming signal as a result of said comparing said second feature vector only with said selected templates.