Method for changing over from a first adaptive data processing version to a second adaptive data processing version

• US 9,009,695 B2
• Filed: 05/09/2007
• Issued: 04/14/2015
• Est. Priority Date: 05/12/2006
• Status: Expired due to Fees

• Current Assignee: Nuance Communications Austria Gmbh (Microsoft Corporation)
• Original Assignee: Nuance Communications Austria Gmbh (Microsoft Corporation)
• Inventors: Unfried, Johannes
• Primary Examiner(s): Zhen, Wei
• Assistant Examiner(s): Bourzik, Brahim
• Application Number:
  US12/300,389
• Publication Number:
  US 20090125899A1
• Time in Patent Office: 2,897 Days
• Field of Search: 717168-178, 704/244
• US Class Current: 717/168
• CPC Class Codes: G06F 8/656 while running
• CPC Class Codes: G10L 15/065 Adaptation
• CPC Class Codes: G10L 15/28 Constructional details of speech recognition systems
• CPC Class Codes: G10L 15/32 Multiple recognisers used in sequence or in parallel; Score combination systems therefor, e.g. voting systems

Abstract

The invention relates to a method and to a system for changing over from a first adaptive data processing version (V1) on data processing means using at least one data model (dm) which is continuously adapted on the basis of data processing results to a second adaptive data processing version (V2) also using at least one data model (DM) to be continuously adapted, characterized in that, in a first phase, the second adaptive data processing version (V2) is used in parallel to the first data processing version (V1), thereby continuously adapting said at least one data model (dm) related to the first version (V1) as well as that data model (DM) related to the second version (V2), and in that the performance of data processing by means of the second version (V2) in checked to comply with a quality criterion, where after in a second phase, as soon as said criterion is met, the results of the data processing by means of the second version (V2) are outputted to be used. The invention further relates to a computer program product having a computer program recorded thereon which is adapted to carry out such a method.

Claims

1. A method for changing from a first adaptive speech recognizer associated with at least one first speech recognition data model to a second adaptive speech recognizer associated with at least one second speech recognition data model, the method comprising using at least one processor to:

execute the first adaptive speech recognizer to process speech data using the at least one first speech recognition data model to provide associated first text, and to adapt the at least one first speech recognition data model based at least in part on the first text;

execute the second adaptive speech recognizer to process the speech data using the at least one second speech recognition data model to generate associated second text, and to adapt the at least one second speech recognition data model based at least in part on the second text, wherein the first and second text generated by the first and second adaptive speech recognizers, respectively, are available for a same use contemporaneous with executing the first adaptive speech recognizer and the second adaptive speech recognizer;

determine whether the performance of the second adaptive speech recognizer complies with at least one quality criterion;

when the performance of the second adaptive speech recognizer does not comply with the at least one quality criterion, output the first text for use; and

when the performance of the second adaptive speech recognizer does comply with the at least one quality criterion, output the second text for use in place of the first text,

wherein the performance of the second adaptive speech recognizer improves as a result of adapting the at least one second speech recognition data model during a period when the first and second adaptive speech recognizers are executed in parallel on the speech data and the first text provided by the first speech recognizer is output for use,

processing the speech data by the second adaptive speech recognizer results in training data used, at least in part, to adapt the at least one second speech recognition data model, and wherein the performance of the second adaptive speech recognizer complies with the at least one quality criterion when the amount of training data exceeds a threshold amount.

2. The method according to claim 1, wherein the performance of the second adaptive speech recognizer complies with the at least one quality criterion when an estimated performance of the second adaptive speech recognizer matches or exceeds an estimated performance of the first adaptive speech recognizer.

3. The method according to claim 1, the method further comprising comparing the second text to the first text, wherein the performance of the second adaptive speech recognizer complies with the at least one quality criterion when the second text is determined to be equal to or better than the first text in accurately representing the speech data.

4. The method according to claim 1, wherein each of the at least one first speech recognition data model and the at least one second speech recognition data model, respectively, comprises at least one model selected from the group consisting of an acoustic model, a language model, a grammar model, and a lexicon model.

13. The method of claim 1, wherein when the performance of the second adaptive speech recognizer does comply with the at least one quality criterion, execution of the second adaptive speech recognizer to process further speech data is continued and execution of the first adaptive speech recognizer is stopped.

14. The method of claim 4, wherein the at least one model selected from the group consisting of the acoustic model, the language model, the grammar model, and the lexicon model comprises a model adapted according to speech from a specific user.

5. At least one non-transitory computer-readable storage device having a computer program recorded thereon which, when executed on at least one processor, performs a method for changing from a first adaptive speech recognizer associated with at least one first speech recognition data model to a second adaptive speech recognizer associated with at least one second speech recognition data model, the method comprising:

executing the first adaptive speech recognizer to process speech data using the at least one first speech recognition data model to provide associated first text and, to adapt the at least one first speech recognition data model based at least in part on the first text;

executing the second adaptive speech recognizer to process the speech data using the at least one second speech recognition data model to generate associated second text, and to adapt the at least one second speech recognition data model based at least in part on the second text, wherein the first and second text generated by the first and second adaptive speech recognizers, respectively, are available for a same use contemporaneous with executing the first adaptive speech recognizer and the second adaptive speech recognizer;

determining whether the performance of the second adaptive speech recognizer complies with at least one quality criterion;

when the performance of the second adaptive speech recognizer does not comply with the at least one quality criterion, outputting the first text for use; and

when the performance of the second adaptive speech recognizer does comply with the at least one quality criterion, outputting the second text for use in place of the first text,

wherein the performance of the second adaptive speech recognizer improves as a result of adapting the at least one second speech recognition data model during a period when the first and second adaptive speech recognizers are executed in parallel on the speech data and the first text provided by the first adaptive speech recognizer is output for use,

wherein processing the speech data by the second adaptive speech recognizer results in training data used, at least in part, to adapt the at least one second speech recognition data model, and wherein the performance of the second adaptive speech recognizer complies with the at least one quality criterion when the amount of training data exceeds a threshold amount.

10. The at least one non-transitory computer-readable storage device of claim 5, wherein the performance of the second adaptive speech recognizer complies with the at least one quality criterion when an estimated performance of the second adaptive speech recognizer matches or exceeds an estimated performance of the first adaptive speech recognizer.

11. The at least one non-transitory computer-readable storage device of claim 5, wherein the performance of the second adaptive speech recognizer complies with the at least one quality criterion when the second text is evaluated as being equal to or better than the first text in accurately representing the speech data.

12. The at least one non-transitory computer-readable storage device of claim 5, wherein each of the at least one first speech recognition data model and the at least one second speech recognition data model respectively comprises at least one model selected from a group consisting of an acoustic model, a language model, a grammar model, and a lexicon model.

15. The at least one non-transitory computer-readable storage device of claim 5, wherein when the performance of the second adaptive speech recognizer does comply with the at least one quality criterion, execution of the second adaptive speech recognizer to process further speech data is continued and execution of the first adaptive speech recognizer is stopped.

16. The at least one non-transitory computer-readable storage device of claim 12, wherein the at least one model selected from the group consisting of the acoustic model, the language model, the grammar model, and the lexicon model comprises a model adapted according to speech from a specific user.

6. A data processing system configured to switch from a first adaptive speech recognizer associated with at least one first speech recognition data model to a second adaptive speech recognizer associated with at least one second speech recognition data model, the system comprising:

at least one storage medium for storing the first adaptive speech recognizer and the second adaptive speech recognizer; and

at least one processor capable of accessing the at least one storage medium, the at least one processor configured to;

execute the first adaptive speech recognizer to process speech data using the at least one first speech recognition data model to provide associated first text and, to adapt the at least one first speech recognition data model based at least in part on the first text;

execute the second adaptive speech recognizer to process the speech data using the at least one second speech recognition data model to generate associated second text, and to adapt the at least one second speech recognition data model based at least in part on the second text, wherein the first and second text generated by the first and second adaptive speech recognizers, respectively, are available for a same use contemporaneous with executing the first adaptive speech recognizer and the second adaptive speech recognizer;

determine whether the performance of the second adaptive speech recognizer complies with at least one quality criterion;

when the performance of the second adaptive speech recognizer does not comply with the at least one quality criterion, output the first text for use; and

when the performance of the second adaptive speech recognizer does comply with the at least one quality criterion, output the second text for use in place of the first text,

wherein the performance of the second adaptive speech recognizer improves as a result of adapting the at least one second speech recognition data model during a period when the first and second adaptive speech recognizers are executed in parallel on the speech data and the first text provided by the first adaptive speech recognizer is output for use,

wherein processing the speech data by the second adaptive speech recognizer results in training data used, at least in part, to adapt the at least one second speech recognition data model, and wherein the performance of the second adaptive speech recognizer complies with the at least one quality criterion when the amount of training data exceeds a threshold amount.

7. The system of claim 6, wherein the performance of the second adaptive speech recognizer complies with the at least one quality criterion when an estimated performance of the second adaptive speech recognizer matches or exceeds an estimated performance of the first adaptive speech recognizer.

8. The system of claim 6, wherein the performance of the second adaptive speech recognizer complies with the at least one quality criterion when the second text is evaluated as being equal to or better than the first text in accurately representing the speech data.

9. The system of claim 6, wherein each of the at least one first speech recognition data model and the at least one second speech recognition data model respectively comprises at least one model selected from the group consisting of an acoustic model, a language model, a grammar model, and a lexicon model.

17. The system of claim 6, wherein when the performance of the second adaptive speech recognizer does comply with the at least one quality criterion, execution of the second adaptive speech recognizer to process further speech data is continued and execution of the first adaptive speech recognizer is stopped.

18. The system of claim 9, wherein the at least one model selected from the group consisting of the acoustic model, the language model, the grammar model, and the lexicon model comprises a model adapted according to speech from a specific user.