RECOGNITION USING RE-RECOGNITION AND STATISTICAL CLASSIFICATION

• US 20100312546A1
• Filed: 06/04/2009
• Published: 12/09/2010
• Est. Priority Date: 06/04/2009
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Levit, Michael, Buntschuh, Bruce, Chang, Shuangyu
• Granted Patent: US 8,930,179 B2
• Time in Patent Office: Days
• US Class Current: 704/9
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G10L 15/08 Speech classification or search
• CPC Class Codes: G10L 15/19 Grammatical context, e.g. disambiguation of the recognition hypotheses based on word sequence rules
• CPC Class Codes: G10L 15/32 Multiple recognisers used in sequence or in parallel; Score combination systems therefor, e.g. voting systems
• CPC Class Codes: G10L 2015/228 of application context

Abstract

Architecture that employs an overall grammar as a set of context-specific grammars for recognition of an input, each responsible for a specific context, such as subtask category, geographic region, etc. The grammars together cover the entire domain. Moreover, multiple recognitions can be run in parallel against the same input, where each recognition uses one or more of the context-specific grammars. The multiple intermediate recognition results from the different recognizer-grammars are reconciled by running re-recognition using a dynamically composed grammar based on the multiple recognition results and potentially other domain knowledge, or selecting the winner using a statistical classifier operating on classification features extracted from the multiple recognition results and other domain knowledge.

Claims

1. A computer-implemented recognition system, comprising:

a constraints component of context-specific constraints for recognition processing of an input into recognition results; and

a reconciliation component for reconciling the recognition results into a single recognition result.

2. The system of claim 1, wherein the constraints include grammars for recognition processing of the input in parallel paths.

3. The system of claim 1, wherein the reconciliation component reconciles the results using re-recognition to generate the single recognition result.

4. The system of claim 3, wherein the re-recognition employs a dynamically composed grammar based on the recognition results.

5. The system of claim 1, wherein the reconciliation component reconciles the results using a statistical classifier that operates on classification features extracted from the recognition results to generate the single recognition result.

6. The system of claim 1, wherein the context-specific constraints include disjointed and intersecting context coverage.

7. The system of claim 1, wherein the recognition processing processes relevant task data to arrive at the single recognition result, the relevant task data includes at least one of recognized strings, utterance level and sub-utterance level confidence scores, speech coverage, relative latencies among concurrent recognitions, prior probabilities of contexts, relative difficulty of each recognition, or consensus among the recognition results.

8. The system of claim 1, further comprising a recognition component for separate recognition processing of the input using a corresponding context-specific constraint in each of parallel paths.

9. The system of claim 1, further comprising a rules component for imposing one or more rules that define determination of the single recognition result.

10. A computer-implemented recognition system, comprising:

a constraints component of context-specific grammars for recognition processing of an utterance input into recognition results;

a recognition component for individual recognition processing of the utterance input in parallel paths using corresponding context-specific grammars; and

a reconciliation component for reconciling the recognition results into a final recognition result.

11. The system of claim 10, wherein the reconciliation component employs a dynamically composed grammar of the recognition results and reconciles the recognition results using re-recognition to generate the final recognition result.

12. The system of claim 10, wherein the reconciliation component reconciles the recognition results using statistical classification that operates on features extracted from the recognition results to generate the final recognition result.

13. The system of claim 10, further comprising a rules component for imposing one or more rules that define determination of the final recognition result and, other domain knowledge that influences features for statistical classification reconciliation and a dynamic grammar for re-recognition reconciliation.

14. The system of claim 10, wherein the reconciliation component reconciles the recognition results by employing statistical analysis prior to re-recognition to determine the final recognition result.

15. A computer-implemented recognition method, comprising:

receiving separate context-specific grammars for processing an utterance input;

recognizing the utterance input in parallel paths using a corresponding context-specific grammar for each path;

generating an intermediate recognition result from each path; and

reconciling the intermediate recognition results into a final recognition result.

16. The method of claim 15, further comprising reconciling the intermediate recognition results using re-recognition of a dynamic grammar generated from the recognition results.

17. The method of claim 15, further comprising:

inputting other domain knowledge during reconciliation of the intermediate recognition results by re-recognition; and

imposing one or more rules to generate the final recognition result.

18. The method of claim 15, further comprising:

inputting other domain knowledge during reconciliation of the intermediate recognition results by statistical classification; and

imposing one or more rules to generate the final recognition result.

19. The method of claim 15, further comprising:

performing statistical analysis; and

assigning relative weights to each path in a re-recognition dynamic grammar.

20. The method of claim 15, further comprising:

waiting for a predetermined amount of time for generation of an intermediate recognition result of a path; and

generating the final recognition result based on intermediate recognition results that are generated within the amount of time.