Dual search acceleration technique for speech recognition

• US 20040148164A1
• Filed: 01/23/2003
• Published: 07/29/2004
• Est. Priority Date: 01/23/2003
• Status: Active Grant

• Current Assignee: Aurilab LLC
• Original Assignee: Aurilab LLC
• Inventors: Baker, James K.
• Granted Patent: US 7,031,915 B2
• Time in Patent Office: Days
• US Class Current: 704/231
• CPC Class Codes: G10L 15/08 Speech classification or search

Abstract

A speech recognition method, system and program product, the method in one embodiment comprising: obtaining input speech data; initiating a first speech recognition search process with at least one hypothesis; initiating a second speech recognition search process with a plurality of hypotheses; obtaining partial results from the second speech recognition search process, where the partial results include an evaluation of at least one hypothesis that the first speech recognition search process has not evaluated at this point in time; and utilizing the partial results to alter the first speech recognition search process.

Claims

1. A speech recognition method, comprising:

obtaining input speech data;

initiating a first speech recognition search process with at least one hypothesis;

initiating a second speech recognition search process with a plurality of hypotheses;

obtaining partial results from the second speech recognition search process, where the partial results include an evaluation of at least one hypothesis that the first speech recognition search process has not evaluated at this point in time; and

utilizing the partial results to alter the first speech recognition search process.

2. The method as defined in claim 1, wherein the first speech recognition search process is a priority queue process.

3. The method as defined in claim 1, wherein the first speech recognition search process is a priority queue process using a best first method.

4. The method as defined in claim 1, wherein the first and second speech recognition search processes are beam search processes.

5. The method as defined in claim 4, wherein the beam search process for the first speech recognition search process has a tighter pruning threshold than the beam search process for the second speech recognition search process.

6. The method as defined in claim 1, wherein the first speech recognition search process proceeds from time frame to time frame at a different rate compared to the second speech recognition search process.

7. A system for speech recognition, comprising:

a component for obtaining input speech data;

a component for performing a first speech recognition search process with at least one hypothesis;

a component for performing a second speech recognition search process with a plurality of hypotheses;

a component for obtaining partial results from the second speech recognition search process, where the partial results include an evaluation of at least one hypothesis that the first speech recognition search process has not evaluated at this point in time; and

a component for utilizing the partial results to alter the first speech recognition search process.

8. A program product for speech recognition, comprising machine-readable program code for causing, when executed, a machine to perform the following method steps:

obtaining input speech data;

initiating a first speech recognition search process with at least one hypothesis;

initiating a second speech recognition search process with a plurality of hypotheses;

obtaining partial results from the second speech recognition search process, where the partial results include an evaluation of at least one hypothesis that the first speech recognition search process has not evaluated at this point in time; and

utilizing the partial results to alter the first speech recognition search process.

9. The program product as defined in claim 8, wherein the first speech recognition search process is a priority queue process.

10. The program product as defined in claim 8, wherein the first speech recognition search process is a priority queue process using a best first method.

11. The program product as defined in claim 8, wherein the first and second speech recognition search processes are beam search processes.

12. The program product as defined in claim 11, wherein the beam search process for the first speech recognition search process has a tighter pruning threshold than the beam search process for the second speech recognition search process.

13. The program product as defined in claim 8, wherein the first speech recognition search process proceeds from time frame to time frame at a different rate compared to the second speech recognition search process.

14. A speech recognition method, comprising:

obtaining input speech data;

initiating a priority queue best first speech recognition search process using a pruning threshold on a best first hypothesis selected from a plurality of hypotheses ranked in an order;

initiating a second speech recognition search process substantially simultaneously on a plurality of hypotheses that includes at least one hypothesis that is not evaluated by the first speech recognition search process;

obtaining partial results at a point in time from the second speech recognition search process, where the partial results include an evaluation of at least one hypothesis that the priority queue search process has not evaluated at the point in time and a score for that hypothesis;

adding a hypothesis to the plurality of hypotheses in the priority queue or reordering the hypotheses in the priority queue based on the partial results;

changing the pruning threshold for the priority queue speech recognition process for each point in time that the second process has evaluated; and

if a criteria is met, selecting a new best first hypothesis and initiating the new priority queue best first speech recognition search process using the new best first search hypothesis at the point in time where the partial results were taken.

15. A speech recognition system, comprising:

a component for obtaining input speech data;

a component for initiating a priority queue best first speech recognition search process using a pruning threshold on a best first hypothesis selected from a plurality of hypotheses ranked in an order;

a component for initiating a second speech recognition search process substantially simultaneously on a plurality of hypotheses that includes at least one hypothesis that is not evaluated by the first speech recognition search process;

a component for obtaining partial results at a point in time from the second speech recognition search process, where the partial results include an evaluation of at least one hypothesis that the priority queue search process has not evaluated at the point in time and a score for that hypothesis;

a component for adding a hypothesis to the plurality of hypotheses in the priority queue or reordering the hypotheses in the priority queue based on the partial results;

a component for changing the pruning threshold for the priority queue speech recognition process for each point in time that the second process has evaluated; and

a component for, if a criteria is met, selecting a new best first hypothesis and initiating the new priority queue best first speech recognition search process using the new best first search hypothesis at the point in time where the partial results were taken.

16. A program product for speech recognition, comprising machine-readable program code for, when executed, causing a machine to perform the following method steps:

obtaining input speech data;

initiating a priority queue best first speech recognition search process using a pruning threshold on a best first hypothesis selected from a plurality of hypotheses ranked in an order;

initiating a second speech recognition search process substantially simultaneously on a plurality of hypotheses that includes at least one hypothesis that is not evaluated by the first speech recognition search process;

obtaining partial results at a point in time from the second speech recognition search process, where the partial results include an evaluation of at least one hypothesis that the priority queue search process has not evaluated at the point in time and a score for that hypothesis;

adding a hypothesis to the plurality of hypotheses in the priority queue or reordering the hypotheses in the priority queue based on the partial results;

changing the pruning threshold for the priority queue speech recognition process for each point in time that the second process has evaluated; and

if a criteria is met, selecting a new best first hypothesis and initiating the new priority queue best first speech recognition search process using the new best first search hypothesis at the point in time where the partial results were taken.

17. A speech recognition method, comprising:

obtaining input speech data;

initiating a first speech recognition search process using a pruning threshold substantially simultaneously on a first plurality of hypothesis;

initiating a second speech recognition search process substantially simultaneously on a second plurality of hypotheses wherein a parameter for the process is different from that parameter in the first speech recognition search process;

obtaining partial results at a point in time from the second speech recognition search process, where the partial results include an evaluation of at least one hypothesis that the priority queue search process has not evaluated at that point in time and a score for that hypothesis;

determining if the score for the at least one hypothesis evaluated by the second speech recognition search process for a particular time frame is better than a best score for that time frame for any hypothesis evaluated by the first speech recognition search process by more than a predetermined amount; and

if yes, then restarting the first speech recognition search process at that point in time on a plurality of hypotheses that include the at least one hypothesis that the priority queue search process has not evaluated and using a new pruning threshold.

18. The method as defined in claim 17, wherein the parameter is a number of hypotheses being evaluated by the second speech recognition search process, which is larger than the number of hypotheses being evaluated by the first speech recognition search process.

19. The method as defined in claim 17, wherein the parameter is that the hypotheses being evaluated by the second speech recognition search process is a superset of the hypotheses being evaluated by the first speech recognition search process.

20. The method as defined in claim 17, wherein the parameter is a pruning margin, and wherein the pruning margin for the first speech recognition search process is tighter than the pruning margin parameter for the second speech recognition search process.

21. A speech recognition system, comprising:

a component for obtaining input speech data;

a component for initiating a first speech recognition search process using a pruning threshold substantially simultaneously on a first plurality of hypothesis;

a component for initiating a second speech recognition search process substantially simultaneously on a second plurality of hypotheses wherein a parameter for the process is different from that parameter in the first speech recognition search process;

a component for obtaining partial results at a point in time from the second speech recognition search process, where the partial results include an evaluation of at least one hypothesis that the priority queue search process has not evaluated at that point in time and a score for that hypothesis;

a component for determining if the score for the at least one hypothesis evaluated by the second speech recognition search process for a particular time frame is better than a best score for that time frame for any hypothesis evaluated by the first speech recognition search process by more than a predetermined amount; and

a component for, if yes, then restarting the first speech recognition search process at that point in time on a plurality of hypotheses that include the at least one hypothesis that the priority queue search process has not evaluated and using a new pruning threshold.

22. The system as defined in claim 21, wherein the parameter is a number of hypotheses being evaluated by the second speech recognition search process, which is larger than the number of hypotheses being evaluated by the first speech recognition search process.

23. The system as defined in claim 21, wherein the parameter is that the hypotheses being evaluated by the second speech recognition search process is a superset of the hypotheses being evaluated by the first speech recognition search process.

24. The system as defined in claim 21, wherein the parameter is a pruning margin, and wherein the pruning margin for the first speech recognition search process is tighter than the pruning margin parameter for the second speech recognition search process.

25. A program product for speech recognition, comprising machine-readable program code for causing, when executed, a machine to perform the following method steps:

obtaining input speech data;

initiating a first speech recognition search process using a pruning threshold substantially simultaneously on a first plurality of hypothesis;

initiating a second speech recognition search process substantially simultaneously on a second plurality of hypotheses wherein a parameter for the process is different from that parameter in the first speech recognition search process;

obtaining partial results at a point in time from the second speech recognition search process, where the partial results include an evaluation of at least one hypothesis that the priority queue search process has not evaluated at that point in time and a score for that hypothesis;

determining if the score for the at least one hypothesis evaluated by the second speech recognition search process for a particular time frame is better than a best score for that time frame for any hypothesis evaluated by the first speech recognition search process by more than a predetermined amount; and

if yes, then restarting the first speech recognition search process at that point in time on a plurality of hypotheses that include the at least one hypothesis that the priority queue search process has not evaluated and using a new pruning threshold.

26. The method as defined in claim 25, wherein the parameter is a number of hypotheses being evaluated by the second speech recognition search process, which is larger than the number of hypotheses being evaluated by the first speech recognition search process.

27. The method as defined in claim 25, wherein the parameter is that the hypotheses being evaluated by the second speech recognition search process is a superset of the hypotheses being evaluated by the first speech recognition search process.

28. The method as defined in claim 25, wherein the parameter is a pruning margin, and wherein the pruning margin for the first speech recognition search process is tighter than the pruning margin parameter for the second speech recognition search process.