METHOD, SYSTEM AND COMPUTER PROGRAM FOR ENHANCED SPEECH RECOGNITION OF DIGITS INPUT STRINGS

• US 20090125306A1
• Filed: 09/19/2008
• Published: 05/14/2009
• Est. Priority Date: 09/19/2007
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: International Business Machines Corporation
• Inventors: Crepy, Hubert, Lejeune, Remi
• Granted Patent: US 8,589,162 B2
• Time in Patent Office: Days
• US Class Current: 704/236
• CPC Class Codes: G10L 15/08 Speech classification or search

Abstract

The present invention proposes a method, system and computer program for speech recognition. According to one embodiment, a method is provided wherein, for an expected input string divided into a plurality of expected string segments, a speech segment is received for each expected string segment. Speech recognition is then performed separately on each said speech segment via the generation, for each said speech segment, of a segment n-best list comprising n highest confidence score results. A global n-best list is then generated corresponding to the expected input string utilizing the segment n-best lists and a final global speech recognition result corresponding to said expected input string is determined via the pruning of the results of the global n-best list utilizing a pruning criterion.

Claims

1. A method for speech recognition comprising:

for an expected input string divided into a plurality of expected string segments, receiving a speech segment for each expected string segment;

performing speech recognition separately on each said speech segment, wherein said performing speech recognition comprises

generating, for each said speech segment, a segment n-best list comprising n highest confidence score results of said speech recognition, where n is an integer;

generating a global n-best list corresponding to said expected input string utilizing said segment n-best lists; and

determining a final global speech recognition result corresponding to said expected input string, wherein said determining said final global speech recognition result comprises

pruning results of said global n-best list utilizing a pruning criterion.

2. The method according to claim 1, wherein receiving a speech segment for each expected string segment further comprises:

receiving a first speech segment corresponding to a first expected string segment from a speaker; and

prompting the speaker to speak a second speech segment corresponding to a second expected string segment.

3. The method according to claim 1, further comprising:

receiving a single continuous speech segment input corresponding to said expected input string from a speaker;

determining one or more time positions within said single continuous speech segment input utilizing a signal received from said speaker to indicate one or more input speech segments; and

dividing said single continuous speech segment input into a plurality of input speech segments utilizing said one or more time positions.

4. The method according to claim 1, wherein performing speech recognition comprises:

performing a grammar analysis speech recognition utilizing a determined maximum length of the expected input string.

5. The method according to claim 1, wherein performing speech recognition comprises:

performing a grammar analysis speech recognition utilizing a determined exact length of the expected input string.

6. The method according to claim 2, further comprising:

receiving data representing a length of at least one expected string segment of said plurality of expected string segments from said speaker; and

restricting, utilizing said data, a grammar analysis of a speech recognition on said at least one expected string segment

7. The method according to claim 1, further comprising:

determining a weight for each result in said global n-best list, wherein said determining said weight comprises

calculating each said weight utilizing a plurality of weights associated with corresponding segment n-best lists results composing said each result in said global n-best list.

8. The method according to claim 7, wherein calculating each said weight comprises:

summing said plurality of weights associated with said corresponding segment n-best lists results.

9. The method according to claim 1, further comprising:

prompting a speaker to repeat a speech segment in response to a determination that speech recognition performed on said speech segment fails to meet a predetermined accuracy threshold.

10. The method according to claim 1, further comprising:

determining an accuracy level of said final global speech recognition result utilizing user input.

11. The method according to claim 1, wherein the expected input string corresponds to a credit card number.

12. The method according to claim 11, wherein the pruning criterion comprises the Luhn algorithm.

13. One or more machine-readable media having stored therein a program product, which when executed by a set of one or more processors causes the set of one or more processors to perform a method comprising:

for an expected input string divided into a plurality of expected string segments, receiving a speech segment for each expected string segment;

performing speech recognition separately on each said speech segment, wherein said performing speech recognition comprises

generating, for each said speech segment, a segment n-best list comprising n highest confidence score results of said speech recognition, where n is an integer;

generating a global n-best list corresponding to said expected input string utilizing said segment n-best lists; and

determining a final global speech recognition result corresponding to said expected input string, wherein said determining said final global speech recognition result comprises

pruning results of said global n-best list utilizing a pruning criterion.

14. A system for speech recognition comprising:

a set of one or more processors;

a memory unit coupled with the set of one or more processors; and

a speech recognition unit operable to,

for an expected input string divided into a plurality of expected string segments, receive a speech segment for each expected string segment;

perform speech recognition separately on each said speech segment, wherein said speech recognition comprises

generating, for each said speech segment, a segment n-best list comprising n highest confidence score results of said speech recognition, where n is an integer;

generate a global n-best list corresponding to said expected input string utilizing said segment n-best lists; and

determine a final global speech recognition result corresponding to said expected input string, wherein determining said final global speech recognition result comprises

pruning results of said global n-best list utilizing a pruning criterion.