Method and apparatus for error correction in a continuous dictation system

• US 5,864,805 A
• Filed: 12/20/1996
• Issued: 01/26/1999
• Est. Priority Date: 12/20/1996
• Status: Expired due to Term

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: International Business Machines Corporation
• Inventors: Olsen, Peder Andreas, Chen, Chengjun Julian, Danis, Catalina Maria, Picheny, Michael Alan, Monkowski, Michael Daniel, Comerford, Liam David, Dharanipragada, Satya
• Primary Examiner(s): Dorvil, Richemond
• Application Number:
  US08/770,390
• Time in Patent Office: 767 Days
• Field of Search: 704/235, 704/275, 704/254, 704/243, 704/244
• US Class Current: 704/235
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue

Abstract

A continuous speech recognition system has the ability to correct errors in strings of words. The error correction method stores data in the system'"'"'s internal state to update probability tables used in developing alternative lists for substitution in misrecognized text.

Claims

1. A method of correcting incorrectly decoded speech in a continuous speech recognition system comprising the steps of:

affording a user a means to select a portion of text representation of the speech as being misrecognized;

displaying a list of a multiplicity of likely alternative sequences of words to the user, wherein the list of alternative sequences is generated by including at least a word preceding said selected portion of text and/or at least a word following said selected portion of text and by decoding an interval of user speech under a constraint that a decoding of either and/or both a preceding or a following word remains unchanged;

affording the user a means to select an alternative sequence;

replacing the selected portion of text representation with a selected alternative sequence.

2. A method a correcting incorrectly decoded speech as in claim 1 wherein the method of affording the user a means to select a portion of text representation of speech includes a step of prompting the user.

3. A method of correcting incorrectly decoded speech as in claim 1 wherein the method of affording the user a means to select an alternative sequence includes a step of prompting the user.

4. A method of correcting correctly decoded speech as in claim 1 wherein said selected portion of misrecognized text representation of speech is a space.

5. A method of correcting incorrectly decoded speech as in claim 1 wherein the list of alternate words is generated including at least a word preceding said selected portion of text as the selected misrecognized portion of text representation of the speech, but displaying only alternative sequences that begin with said words preceding said selected portion.

6. A method of correcting incorrectly decoded speech as in claim 1 wherein the list of alternate words is generated including at least a word following said selected portion of text as the selected misrecognized portion of text representation of the speech, but displaying only alternative sequences that end with said words following said selected portion.

7. A method of correcting incorrectly decoded speech as in claim 1 wherein the list of alternate words is generated including at least a word preceding said selected portion of text and at least a word following said selected portion of text as the selected misrecognized portion of text representation of the speech, but displaying only alternative sequences that begin with said words preceding said selected portion and end with said words following said selected portion.

8. A method of correcting incorrectly decoded speech as in claim 1 further comprising providing a playback of speech corresponding to a portion of the text representation of the speech as being corrected.

9. A method of correcting incorrectly decoded speech as in claim 8 wherein said playback of speech includes at least a word preceding said portion of text representation.

10. A method of correcting incorrectly decoded speech as in claim 8 wherein said playback of speech includes at least a word following said portion of text representation.

11. A method of correcting incorrectly decoded speech as in claim 8 wherein said playback of speech includes at least a word preceding said portion of text representation and at least a word following said portion of text representation.

12. A method of correcting misrecognized speech continuous speech recognition system comprising the steps of:

a) affording a user a means to select a portion of a text representation of speech as being misrecognized;

b) displaying a list of a multiplicity of likely alternative sequences of words to the user, wherein the list of alternative sequences is generated by including as least a word preceding said selected portion of text and/or at least a word following said selected portion of text and by decoding an interval of user speech under a constraint that a decoding of either and/or both a preceding or a following word remains uncharged;

c) affording the user a means to select a portion of an alternative sequence;

d) using a starting and a ending boundary of a selected portion of an alternative sequence to identify a portion of a selected portion of a text representation to be replaced;

e) replacing said portion of the selected portion of a representation with a selected portion of an alternative sequence;

f) identifying any remaining misrecognized speech;

g) displaying a new list of alternative sequences for identified remaining misrecognized speech;

h) affording the user a means to select an alternative sequence from said new list; and

i) replacing said remaining misrecognized with a selected alternative sequence from the new list.

13. A method of correcting misrecognized speech as in claim 12 further comprising steps of repeating said steps of f) to i) until said text representation of misrecognized speech is completely replaced.

14. A method of correcting misrecognized speech in a continuous speech recognition system as in claim 12 in which a partial word may be selected.

15. A method of correcting misrecognized speech in a continuous speech recognition system as in claim 12 wherein said partially selected alternative sequence has an ending time which falls within a word in said user selected text and further comprising the step of decoding for acoustic data representing a remainder of said word is displayed to the user with at least one place holder symbol, following replacing said portion of the selected portion of a text representation.

16. A method of correcting misrecognized speech in a continuous speech recognition system as in claim 15 wherein said place holder symbol is selected from the group consisting of graphical symbols and alphanumeric characters.

17. A method of correcting misrecognized speech in a continuous speech recognition system as in claim 12 wherein said step of affording a user means to select a portion of a text representation of speech includes at least a step of prompting the user.

18. A method of correcting misrecognized speech in a continuous speech recognition system as in claim 12 wherein said step of affording a user means to select a portion of an alternative sequence includes at least a step of prompting the user.

19. A method of correcting misrecognized speech in a continuous speech recognition system as in claim 12 wherein said step of affording a user means to select an alternative sequence from said new list includes at least a step of prompting the user.

20. A method of correcting misrecognized speech in a continuous speech recognition system as in claim 12, wherein said step of identifying any remaining misrecognized speech comprises the steps of:

determining a beginning time and an ending time of a portion of an alternative sequence;

determining a beginning time and an ending time of a user selected portion of text;

replacing a portion of said user selected text with said user selected portion of an alternative sequence; and

changing from said beginning and ending times of said user selected portion of text, to reflect removing said alternative sequence replacing a portion of said user selection.

21. The method of correcting misrecognized speech in continuous speech recognition system as in claim 12, wherein said list of a multiplicity of likely alternative sequences of words is ordered by a method comprising the steps of:

matching a beginning time of each alternative sequence of words with a beginning time of said selected portion of text;

matching an ending time of each alternative sequence of words with an ending time of said selected portion of text; and

ranking said alternative sequences of words according to how closely the beginning and ending times of the alternative sequences of words match the beginning and ending times of said selected portion of text.

22. The method of correcting misrecognized speech in continuous speech recognition system as in claim 12, wherein said list of a multiplicity of likely alternative sequences of words is ordered by a method comprising the steps of:

determining acoustic data for said selected portion of text;

determining acoustic data for said alternative sequence of words;

matching each alternative sequence of words by acoustic data to said selected portion of text; and

ranking said alternative sequences of words according to how closely acoustic data of each alternative sequence matches acoustic data of said selected portion of text.

23. The method of correcting misrecognized speech in continuous speech recognition system as in claim 12, wherein said list of a multiplicity of likely alternative sequences of words is ordered by a method comprising the steps of:

determining language and acoustic data for said selected portion of text;

determining language and acoustic data for said alternative sequence of words;

matching each alternative sequence of words by language acoustic data to said selected portion of text; and

ranking said alternative sequences of words according to how closely language acoustic data of each alternative sequence matches language and acoustic data of said selected portion of text.

24. A method as in claim 12 wherein said list of a multiplicity of alternative word sequences is reduced comprising the steps of:

grouping each of said alternative word sequences which have the same decoded words for any given portion of the sequence; and

displaying at least the most likely alternatives which have different decoded word patterns.