System and method of spoken language understanding using word confusion networks

1. A computer-implemented method, comprising:

converting in a processor a word lattice that describes multiple hypotheses of a received input utterance at an automatic speech recognition (ASR) engine into a modified word lattice, wherein transitions without any input in the word lattice are represented as epsilon transitions in the modified word lattice; and

performing in a processor spoken language understanding for the received input utterance based on the modified word lattice.

2. The computer-implemented method of claim 1 , wherein the converting further comprises aligning words in the word lattice.

3. The computer-implemented method of claim 2 , wherein the converting further comprises converting the word lattice into a word confusion network.

4. The computer-implemented method of claim 1 , wherein the converting is performed at an ASR engine associated with said processor and performing spoken language understanding occurs at a spoken language understanding engine associated with said processor.

5. The computer-implemented method of claim 1 , wherein the performing further comprises:

calculating an intersection of the modified word lattice and a finite state machine including a plurality of salient grammar fragments; and

determining a classification type for the input utterance based on the calculated intersection.

6. The computer-implemented method of claim 5 , further comprising preprocessing in said processor the modified word lattice, wherein the calculating comprises calculating an intersection of the preprocessed modified word lattice and the finite state machine.

7. The computer-implemented method of claim 5 , wherein the calculating comprises calculating the intersection of the modified word lattice and a finite state machine that consists of all of the salient grammar fragments.

8. The computer-implemented method of claim 5 , wherein the calculating comprises calculating an intersection of the modified word lattice and the finite state machine including the plurality of salient grammar fragments to produce a set of salient phrases.

9. The computer-implemented method of claim 8 , further comprising filtering and parsing in said processor the set of salient phrases.

10. A module in an automatic speech recognition (ASR) engine that is configured to control a processor to:

determine an intersection of a word lattice that converts a word lattice that describes multiple hypotheses of a received input utterance into a modified word lattice, wherein transitions without any input in the word lattice are represented as epsilon transitions in the modified word lattice; and

a module configured to perform spoken language understanding for the received input utterance based on the modified word lattice.

11. The speech processing system using the module of claim 10 , wherein the module configured to convert further aligns words in the word lattice.

12. The speech processing system using the module of claim 11 , wherein the module configured to convert further converts the word lattice into a word confusion network.

13. The speech processing system using the module of claim 12 , wherein posterior probabilities of the word confusion network are used as confidence scores.

14. The speech processing system using the module of claim 10 , wherein the module configured to convert conform further calculates an intersection of the modified word lattice and a finite state machine including a plurality of salient grammar fragments and determines a classification type for the input utterance based on the calculated intersection.

15. The speech processing system using the module of claim 14 , further comprising a module configured to preprocess the modified word lattice, wherein the module configured to calculate further calculates an intersection of the preprocessed modified word lattice and the finite state machine.

16. The speech processing system using the module of claim 14 , wherein the module configured to calculate further calculates the intersection of the modified word lattice and a finite state machine that consists of all of the salient grammar fragments.

17. The speech processing system using the module of claim 14 , wherein the module configured to calculate further calculates an intersection of the modified word lattice and the finite state machine including the plurality of salient grammar fragments to produce a set of salient phrases.

18. The speech processing system using the module of claim 17 , further comprising a module configured to filter and parse the set of salient phrases.

19. The tangible computer-readable medium that stores a program for controlling a computing device to perform a speech processing method, the method comprising:

converting a word lattice that describes multiple hypotheses of a received input utterance at an automatic speech recognition (ASR) engine into a modified word lattice, wherein transitions without any input in the word lattice are represented as epsilon transitions in the modified word lattice; and

performing spoken language understanding for the received input utterance based on the modified word lattice.

20. The tangible computer-readable medium of claim 19 , wherein the performing spoken language understanding further comprises:

calculating an intersection of the modified word lattice and a finite state machine including a plurality of salient grammar fragments; and

determining a classification type for the input utterance based on the calculated intersection.

21. The tangible computer-readable medium of claim 19 , wherein the converting further comprises converting the word lattice into a word confusion network wherein posterior probabilities are used as confidence scores.