Method and apparatus for error correction in speech recognition applications

• US 7,756,710 B2
• Filed: 07/13/2006
• Issued: 07/13/2010
• Est. Priority Date: 07/13/2006
• Status: Active Grant

• Current Assignee: SRI International, Inc.
• Original Assignee: SRI International, Inc.
• Inventors: Cesari, Federico, Franco, Horacio, Zheng, Jing, Myers, Gregory, Cowan, Cregg
• Primary Examiner(s): MCFADDEN, SUSAN IRIS
• Application Number:
  US11/486,266
• Publication Number:
  US 20100125458A1
• Time in Patent Office: 1,461 Days
• Field of Search: 704/255
• US Class Current: 704/255
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue
• CPC Class Codes: G10L 2015/223 Execution procedure of a spoken command

Abstract

In one embodiment, the present invention is a method and apparatus for error correction in speech recognition applications. In one embodiment, a method for recognizing user speech includes receiving a first utterance from the user, receiving a subsequent utterance from the user, and combining acoustic evidence from the first utterance with acoustic evidence from the subsequent utterance in order to recognize the first utterance. It is assumed that, if the first utterance has been incorrectly recognized on a first attempt, the user will repeat the first utterance (or at least the incorrectly recognized portion of the first utterance) in the subsequent utterance.

Claims

1. A method for recognizing user speech, the method comprising:

receiving a first utterance from said user;

receiving a subsequent utterance from said user, wherein said subsequent utterance comprises at least one pre-defined keyword that indicates that acoustic evidence from said first utterance should be combined with acoustic evidence from said subsequent utterance; and

combining said acoustic evidence from said first utterance with said acoustic evidence from said subsequent utterance in order to recognize said first utterance,

wherein at least one of said;

receiving said first utterance, said receiving said subsequent utterance, or said combining is performed using a processor.

2. The method of claim 1, wherein said subsequent utterance comprises at least a portion of said first utterance.

3. The method of claim 1, further comprising:

providing a recognition result produced for said first utterance to said user.

4. The method of claim 1, wherein said at least one pre-defined keyword is replaced with a null word in a word lattice associated with said subsequent utterance.

5. The method of claim 1, wherein said acoustic evidence from said first utterance comprises a first word lattice and said acoustic evidence from said subsequent utterance comprises a subsequent word lattice, each of said first word lattice and said subsequent word lattice comprising:

a plurality of lattice positions, each of said plurality of lattice positions corresponding to a temporal interval in said first utterance or in said subsequent utterance;

at each of said plurality of lattice positions, one or more words that are believed to have been uttered during the temporal interval; and

at each of said one or more words, a probability score that indicates a likelihood that an associated one of the one or more words was uttered during the temporal interval.

6. The method of claim 5, wherein said combining comprises:

aligning said first word lattice with said subsequent word lattice to form an aligned word lattice; and

selecting at least one phrase from said aligned word lattice as a recognition result corresponding to said first utterance.

7. The method of claim 6, wherein said aligned word lattice comprises:

a plurality of aligned lattice positions, each of the plurality of aligned lattice positions corresponding to a given lattice position in at least one of;

said first word lattice or said subsequent word lattice;

at each of said plurality of aligned lattice positions, at least one word, said at least one word comprising;

each of said one or more words associated with the given lattice position in the first word lattice and each of said one or more words associated with the given lattice position in the subsequent word lattice; and

at said at least one word, a combined probability score that comprises a combination of;

the probability score associated with the at least one word at the given lattice position in the first word lattice and the probability score associated with the at least one word at the given lattice position in the subsequent word lattice.

8. The method of claim 7, wherein said selecting comprises:

identifying at least one phrase in said aligned word lattice having a highest combined probability score.

9. The method of claim 8, wherein said identifying comprises:

identifying, at each of said plurality of aligned lattice positions, an associated word having a highest combined probability score; and

concatenating words identified at each of said plurality of aligned lattice positions to form said at least one phrase.

10. The method of claim 7, further comprising:

identifying, in at least one of said first word lattice and said subsequent word lattice, one or more deletions at one or more of said plurality of lattice positions; and

removing deletions associated with probability scores that exceed a predefined probability threshold.

11. An apparatus for recognizing user speech, comprising:

means for receiving a first utterance from a user;

means for receiving a subsequent utterance from said user, wherein said subsequent utterance comprises at least one pre-defined keyword that indicates that acoustic evidence from said first utterance should be combined with acoustic evidence from said subsequent utterance; and

means for combining said acoustic evidence from said first utterance with said acoustic evidence from said subsequent utterance in order to recognize said first utterance.

12. A computer readable medium containing an executable program for recognizing user speech, where the program performs steps of:

receiving a first utterance from said user;

receiving a subsequent utterance from said user, wherein said subsequent utterance comprises at least one pre-defined keyword that indicates that acoustic evidence from said first utterance should be combined with acoustic evidence from said subsequent utterance; and

combining said acoustic evidence from said first utterance with said acoustic evidence from said subsequent utterance in order to recognize said first utterance.

13. The computer readable medium of claim 12, wherein said subsequent utterance comprises at least a portion of said first utterance.

14. The computer readable medium of claim 12, further comprising:

providing a recognition result produced for said first utterance to said user.

15. The computer readable medium of claim 12, wherein said at least one pre-defined keyword is replaced with a null word in a word lattice associated with said subsequent utterance.

16. The computer readable medium of claim 12, wherein said acoustic evidence from said first utterance comprises a first word lattice and said acoustic evidence from said subsequent utterance comprises a subsequent word lattice, each of said first word lattice and said subsequent word lattice comprising:

a plurality of lattice positions, each of said plurality of lattice position positions corresponding to a temporal interval in said first utterance or in said subsequent utterance;

at each of said plurality of lattice positions, one or more words that are believed to have been uttered during the temporal interval; and

at each of said one or more words, a probability score that indicates a likelihood that an associated one of the one or more words was uttered during the temporal interval.

17. The computer readable medium of claim 16, wherein said combining comprises:

aligning said first word lattice with said subsequent word lattice to form an aligned word lattice; and

selecting at least one phrase from said aligned word lattice as a recognition result corresponding to said first utterance.

18. The computer readable medium of claim 17, wherein said aligned word lattice comprises:

a plurality of aligned lattice positions, each of the plurality of aligned lattice positions corresponding to a given lattice position in at least one of;

said first word lattice or said subsequent word lattice;

at each of said plurality of aligned lattice positions, at least one word, said at least one word comprising;

each of said one or more words associated with the given lattice position in the first word lattice and each of said one or more words associated with the given lattice position in the subsequent word lattice; and

at said at least one word, a combined probability score that comprises a combination of;

the probability score associated with the at least one word at the given lattice position in the first word lattice and the probability score associated with the at least one word at the given lattice position in the subsequent word lattice.

19. The computer readable medium of claim 18, wherein said selecting comprises:

identifying, at each of said plurality of aligned lattice positions, an associated word having a highest combined probability score; and

concatenating words identified at each of said plurality of aligned lattice positions to form said at least one phrase.

20. The computer readable medium of claim 18, further comprising:

identifying, in at least one of said first word lattice and said subsequent word lattice, one or more deletions at one or more of said plurality of lattice positions; and

removing deletions associated with probability scores that exceed a predefined probability threshold.