METHOD AND SYSTEM FOR IMPROVED SPEECH RECOGNITION

• US 20090287483A1
• Filed: 05/14/2008
• Published: 11/19/2009
• Est. Priority Date: 05/14/2008
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: International Business Machines Corporation
• Inventors: Co, Raymond L., Jan, Ee-ee, Lubensky, David M.
• Granted Patent: US 7,680,661 B2
• Time in Patent Office: Days
• US Class Current: 704/231
• CPC Class Codes: G10L 15/08 Speech classification or search
• 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

A method for speech recognition includes: prompting a user with a first query to input speech into a speech recognition engine; determining if the inputted speech is correctly recognized; wherein in the event the inputted speech is correctly recognized proceeding to a new task; wherein in the event the inputted speech is not correctly recognized, prompting the user repeatedly with the first query to input speech into the speech recognition engine, and determining if the inputted speech is correctly recognized until a predefined limit on repetitions has been met; wherein in the event the predefined limit has been met without correctly recognizing the inputted user speech, prompting speech input from the user with a secondary query for redundant information; and cross-referencing the user'"'"'s n-best result from the first query with the n-best result from the second query to obtain a top hypothesis.

Claims

1. A method for improved speech recognition utilizing redundant information, the method comprising:

prompting a user with a first query for first information as input speech into a speech recognition engine;

receiving the user'"'"'s inputted speech with the first information;

determining if the inputted speech with the first information is correctly recognized by the speech recognition engine;

wherein in the event the first information from the inputted speech is correctly recognized proceeding to a new task;

in response to a determination that the first information is not correctly recognized,

prompting speech input from the user with a secondary query for secondary information;

wherein the secondary information can be used to identify the first information;

cross-referencing the user'"'"'s n-best result from the first query with the n-best result from the second query to obtain a top hypothesis; and

wherein in the event the top hypothesis is incorrect the user is transferred to a human agent.

3. The method of claim 1, wherein the cross-referencing is carried out with a join n-best algorithm;

and

wherein the join n-best algorithm comprises at least one of the following;

best join scores, weighted join scores, or fusion scores.

5. The method of claim 1, wherein in the event the top hypothesis is incorrect the method further comprises:

querying the user for one or more additional pieces of secondary information;

generating a new top hypothesis by cross-referencing a set of n-best results from the additional secondary queries; and

wherein in the event the number of secondary queries reaches a preset maximum number of queries, without correctly recognizing the user'"'"'s inputted speech, the user is transferred to a human agent.

6. The method of claim 1, wherein:

in the event the inputted speech for the first information is not correctly recognized, prompting the user repeatedly with the first query to input speech for the first information into the speech recognition engine, and determining if the inputted speech with the first information is correctly recognized until a predefined limit on repetitions has been met; and

in the event the predefined limit has been met without correctly recognizing the first information from the inputted user speech, then prompting speech input from the user with a secondary query for secondary information.

2. The method of claim I, wherein based on the cross-referencing the user is provided recognized speech matches to the first query in ranked order.

4. The method of claim 64 wherein the method further comprises repeatedly asking the second query until at least one of the following occurs:

the secondary information is correctly recognized, a preset threshold of a false accept tolerance is exceeded, and a preset threshold of maximum number of secondary queries is exceeded.