Pattern recognition

• US 20040039573A1
• Filed: 03/27/2003
• Published: 02/26/2004
• Est. Priority Date: 03/27/2002
• Status: Active Grant

• Current Assignee: WSOU Investments, LLC (WSOU Holdings, LLC)
• Original Assignee: Nokia Corporation
• Inventors: Vasilache, Marcel
• Granted Patent: US 7,912,715 B2
• Time in Patent Office: Days
• US Class Current: 704/242
• 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 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, comprising:

comparing a first feature vector in said sequence 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, and

comparing a 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 number of successive feature vectors is determined in response to a control signal.

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

17. A method according to claim 1, wherein said templates are Gaussian mixture densities of Hidden Markov Models (HMMs).

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

19. A method according to claim 17, wherein said pattern recognition includes computing a state likelihood for a HMM with respect to a feature vector.

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

21. A computer program product, comprising computer program code portions arranged to, when executed by a computer processor, determine a set of distortion measures in a pattern recognition process by performing the steps of:

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 in said sequence 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, and

comparing a second feature vector only with said selected templates.

22. A computer program product according to claim 21, stored on a computer readable medium.

23. A device 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, comprising:

means for comparing a first feature vector in said sequence 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.

24. A device according to claim 23, wherein said selected templates include the templates resulting in the lowest distortion measures when compared to said first feature vector.

25. A device according to claim 23, further comprising means for including distortion measures computed with respect to a different feature vector, in said set of distortion measures.

26. A device according to claim 23, wherein said means for comparing said second feature vector are arranged to compare a number of successive feature vectors only with said selected templates.

27. A device according to claim 23, further comprising a control module, adapted to detect the processor load and to adjust the number of successive feature vectors in response to a said load.

28. A speech recognizer comprising a device according to claim 23.

29. A communication device comprising a speech recognizer according to claim 28.

30. A system for pattern recognition, comprising:

means for forming a sequence of feature vectors from a digitized incoming signal,

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

means for comparing a first feature vector in said sequence 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.

31. A system according to claim 30, 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.