Speech recognition rejection method using generalized additive models

• US 6,006,182 A
• Filed: 09/22/1997
• Issued: 12/21/1999
• Est. Priority Date: 09/22/1997
• Status: Expired due to Term

• Current Assignee: RPX Clearinghouse LLC (RPX Corporation)
• Original Assignee: Northern Telecom Limited (Nortel Networks Corporation)
• Inventors: Crespo, Jean-Francois, Sabourin, Michael, Tremblay, Real, Gupta, Vishwa, Robillard, Serge, Fakhr, Waleed
• Primary Examiner(s): MCFADDEN, SUSAN IRIS
• Application Number:
  US08/934,892
• Time in Patent Office: 820 Days
• Field of Search: 704/231, 704/233, 704/234, 704/238, 704/239, 704/240, 704/251, 704/256
• US Class Current: 704/231
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]

Abstract

Systems and methods consistent with the present invention determine whether to accept one of a plurality of intermediate recognition results output by a speech recognition system as a final recognition result. The system first combines a plurality of speech rejection features into a feature function in which weights are assigned to each rejection feature in accordance with a recognition accuracy of each rejection feature. Feature values are then calculated for each of the rejection features using the plurality of intermediate recognition results. The system next computes the feature function according to the calculated feature values to determine a rejection decision value. Finally, one of the plurality of intermediate recognition results is accepted as the final recognition result according to the rejection decision value.

Claims

1. A computer-implemented method for accepting one of a plurality of intermediate recognition results output by a speech recognition system as a final recognition result, the method comprising the steps of:

combining a plurality of speech rejection features into a generalized additive model (GAM) feature function in which weights are assigned to each rejection feature in accordance with a recognition accuracy of each rejection feature;

calculating feature values for each of the rejection features using the plurality of intermediate recognition results;

computing the GAM feature function according to the calculated feature values to determine a rejection decision value; and

accepting one of the plurality of intermediate recognition results as the final recognition result according to the rejection decision value.

2. The method of claim 1, wherein the GAM feature function includes a plurality of coefficients, and wherein the method further includes the step of:

using a speech training set to determine the coefficients of the GAM feature function.

3. The method of claim 1, wherein the feature function is defined by the following equation:

##EQU2## where;

θ

represents an offset value;

δ

represents a weight;

F(f,α

) represents a function describing each rejection feature f;

J represents the number of rejection features that are combined in the equation for R(f);

K represents the order of function F; and

R(f) ranges in value from 0 to 1.

4. The method of claim 3, wherein the function F(f_j,α

) describing each rejection feature is defined by the following equation;

space="preserve" listing-type="equation">F(f.sub.i,α

)=α

.sub.i.1 f.sub.i +α

.sub.i.2 f.sub.i.sup.2 +α

.sub.i.3 f.sub.i.sup.3.

5. The method of claim 3, wherein the function F(f_j,α

) is a polynomial having an order K and wherein the combining step further includes the substep of;

measuring a modeling error produced by the feature function for a plurality of values for K; and

selecting the order K according to the measured modeling error.

6. The method of claim 3, wherein the a coefficients for each rejection feature are determined using a gradient logistic regression method.

7. The method of claim 1, wherein the rejection features are any one of the following rejection features:

(1) a score-per-frame feature;

(2) a score ratio feature;

(3) a score spread feature;

(4) a normalized score difference feature;

(5) a allophone number feature;

(6) a confusability feature; and

(7) the weighted Levinstein distance between words.

8. The method of claim 1, wherein the step of accepting one of the plurality of intermediate recognition results as the final recognition result further includes the substeps of:

comparing the rejection decision value to a threshold value; and

accepting one of the plurality of intermediate recognition results as the final recognition result according to whether the rejection decision value is above the threshold value.

9. The method of claim 1, wherein the plurality of intermediate recognition results are ranked according to a recognition score associated with each intermediate recognition result, and wherein the step of accepting one of the plurality of intermediate recognition results as the final recognition result further includes the substep of:

accepting the top ranked intermediate recognition result as the final recognition result.

10. A system for accepting one of a plurality of intermediate recognition results output by a speech recognition system as a final recognition result, the system comprising:

means for combining a plurality of speech rejection features into a generalized additive model (GAM) feature function in which weights are assigned to each rejection feature in accordance with a recognition accuracy of each rejection feature;

means for calculating feature values for each of the rejection features using the plurality of intermediate recognition results;

means for computing the GAM feature function according to the calculated feature values to determine a rejection decision value; and

means for accepting one of the plurality of intermediate recognition results as the final recognition result according to the rejection decision value.

11. The system of claim 10, wherein coefficients of the GAM feature function are predetermined.

12. The system of claim 10, wherein the feature function is defined by the following equation:

##EQU3## where;

θ

represents an offset value;

δ

represents a weight;

F(f,α

) represents a function describing each rejection feature f;

J represents the number of rejection features that are combined in the equation for R(f);

K represents the order of function F; and

R(f) ranges in value from 0 to 1.

13. The system of claim 12, wherein the function F(f_j,α

) describing each rejection feature is defined by the following equation;

space="preserve" listing-type="equation">F(f.sub.i,α

)=α

.sub.i.1 f.sub.i +α

.sub.i.2 f.sub.i.sup.2 +α

.sub.i.3 f.sub.i.sup.3.

14. The system of claim 12, wherein the function F(f_j,α

) is a polynomial having an order K and wherein the means for combining further includes;

means for measuring a modeling error produced by the feature function for a plurality of values for K; and

means for selecting the order K according to the measured modeling error.

15. The system of claim 12, wherein the means for combining further includes:

means for determining the a coefficients for each rejection feature using a gradient logistic regression method.

16. The system of claim 10, wherein the rejection features are any one of the following rejection features:

(1) a score-per-frame feature;

(2) a score ratio feature;

(3) a score spread feature;

(4) a normalized score difference feature;

(5) a allophone number feature;

(6) a confusability feature; and

(7) the weighted Levinstein distance between words.

17. The system of claim 10, wherein the means for accepting one of the plurality of intermediate recognition results as the final recognition result further includes:

means for comparing the rejection decision value to a threshold value; and

means for accepting one of the plurality of intermediate recognition results as the final recognition result according to whether the rejection decision value is above the threshold value.

18. The system of claim 10, wherein the plurality of intermediate recognition results are ranked according to a recognition score associated with each intermediate recognition result, and wherein the means for accepting one of the plurality of intermediate recognition results as the final recognition result further includes:

means for accepting the top ranked intermediate recognition result as the final recognition result.

19. The system of claim 10, wherein the speech recognition system is a voice activated dialing system.

20. A computer program product containing instructions for causing a computer to execute a method for accepting one of a plurality of intermediate recognition results output by a speech recognition system as a final recognition result, the computer program product comprising:

instructions for combining a plurality of speech rejection features into a generalized additive model (GAM) feature function in which weights are assigned to each rejection feature in accordance with a recognition accuracy of each rejection feature;

instructions for calculating feature values for each of the rejection features using the plurality of intermediate recognition results;

instructions for computing the GAM feature function according to the calculated feature values to determine a rejection decision value; and

instructions for accepting one of the plurality of intermediate recognition results as the final recognition result according to the rejection decision value.

21. The computer program product of claim 20, wherein the feature function is defined by the following equation:

##EQU4## where;

θ

represents an offset value;

δ

represents a weight;

F(f,α

) represents a function describing each rejection feature f;

J represents the number of rejection features that are combined in the equation for R(f);

K represents the order of function F; and

R(f) ranges in value from 0 to 1.

22. The computer program product of claim 21, wherein the function F(f_j,α

) describing each rejection feature is defined by the following equation;

space="preserve" listing-type="equation">F(f.sub.j,α

)=α

.sub.j,1 f.sub.j +α

.sub.j,2 f.sub.j.sup.2 +α

.sub.j,3 f.sub.j.sup.3.

23. The computer program product of claim 21, wherein the function F(f_j,α

) is a polynomial having an order K and wherein the combining instructions further includes;

instructions for measuring a modeling error produced by the feature function for a plurality of values for K; and

instructions for selecting the order K according to the measured modeling error.

24. The computer program product of claim 21, wherein the a coefficients for each rejection feature are determined using a gradient logistic regression method.

25. The computer program product of claim 20, wherein the rejection features are any one of the following rejection features:

(1) a score-per-frame feature;

(2) a score ratio feature;

(3) a score spread feature;

(4) a normalized score difference feature;

(5) a allophone number feature;

(6) a confusability feature; and

(7) the weighted Levinstein distance between words.

26. The computer program product of claim 20, wherein the instructions for accepting one of the plurality of intermediate recognition results as the final recognition result further includes:

instructions for comparing the rejection decision value to a threshold value; and

instructions for accepting one of the plurality of intermediate recognition results as the final recognition result according to whether the rejection decision value is above the threshold value.

27. The computer program product of claim 20, wherein the plurality of intermediate recognition results are ranked according to a recognition score associated with each intermediate recognition result, and wherein the instructions for accepting one of the plurality of intermediate recognition results as the final recognition result further includes:

instructions for accepting the top ranked intermediate recognition result as the final recognition result.

28. The computer program product of claim 20, wherein coefficients of the GAM feature function are predetermined.