Determining distortion measures in a pattern recognition process

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.