Speech recognition using re-utterance recognition

• US 20050159950A1
• Filed: 12/05/2004
• Published: 07/21/2005
• Est. Priority Date: 09/05/2001
• Status: Active Grant

• Current Assignee: Cerence Operating Company (Cerence Inc.)
• Original Assignee: Voice Signal Technologies Incorporated (Microsoft Corporation)
• Inventors: Roth, Daniel L., Cohen, Jordan R.
• Granted Patent: US 7,444,286 B2
• Time in Patent Office: Days
• US Class Current: 704/236
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue

Abstract

The present invention relates to speech recognition that enables a user to perform re-utterance recognition, in which speech recognition is performed upon both a second saying of a sequence of one or more words and upon an earlier saying of the same sequence to help the speech recognition better select one or more best scoring text sequences for the utterances.

Claims

1. A method of speech recognition comprising:

receiving an original utterance of one or more words;

performing an original speech recognition upon the original utterance;

producing a user perceivable output representing one or more sequences of one or more words selected by the recognition as most likely corresponding to the utterance;

providing a user interface that allows a user to select to perform a re-utterance recognition upon a part of the original utterance corresponding to all or a selected part of the user perceivable output; and

responding to a user selection to perform a re-utterance recognition upon all or a part of the original utterance by;

treating a second utterance received in association with the selection as a re-utterance of the selected portion of the original utterance; and

performing speech recognition upon the re-utterance to select one or more sequences of one or more words considered to most likely match the re-utterance based on the scoring of the one or more words against both the re-utterance and the selected portion of the original utterance.

2. A method as in claim 1 wherein:

the original recognition of the original utterance is by continuous speech recognition; and

the re-utterance is recognized by discrete speech recognition.

3. A method as in claim 2 wherein the number of utterances detected with a re-utterance recognized by discrete recognition is used to determine the number of words allowable in sequences of one or more words recognized for the original utterance after the re-utterance.

4. A method as in claim 1 wherein both the original utterance and the re-utterance are recognized by discrete speech recognition.

5. A method as in claim 1 wherein both the original utterance and the re-utterance are recognized by continuous speech recognition.

6. A method as in claim 1 wherein the selection of a sequences of one or more words considered to most likely match both the re-utterance and the selected portion of the original utterance is used to update acoustic models with data from the selected portion of the original utterance.

7. A method as in claim 1 wherein:

the user interface allows a user to select one or more word filtering inputs, each indicating that the desired output has certain characteristics, to be used in conjunction with the re-utterance recognition; and

the process of selecting of one or more sequences as most likely matching both the re-utterance and the original utterance also uses the selected filtering inputs to favor the selection of any recognition candidates having the selected characteristics.

8. A method as in claim 7 wherein the user interface allows a user to select alphabetic filtering inputs indicating that the desired output contains a word containing a sequence of one or more specified letters.