CONFIDENCE MEASURE SYSTEM USING A NEAR-MISS PATTERN

• US 20010018654A1
• Filed: 11/13/1998
• Published: 08/30/2001
• Est. Priority Date: 11/13/1998
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: HON, HSIAO-WUEN, GUNAWARDANA, ASELA J.R.
• Granted Patent: US 6,571,210 B2
• Time in Patent Office: Days
• US Class Current: 704/257
• CPC Class Codes: G10L 15/08 Speech classification or search

Abstract

A method and system of performing confidence measure in a speech recognition system includes receiving an utterance of input speech and creating a near-miss pattern or a near-miss list of possible word entries for the utterance. Each word entry includes an associated value of probability that the utterance corresponds to the word entry. The near-miss list of possible word entries is compared with corresponding stored near-miss confidence templates. Each word in the vocabulary (or keyword list) of near-miss confidence template, which includes a list of word entries and each word entry in each list includes an associated value. Confidence measure for a particular hypothesis word is performed based on the comparison of the values in the near-miss list of possible word entries with the values of the corresponding near-miss confidence template.

Claims

1. A method of performing speech recognition, the method comprising:

receiving an utterance of input speech;

creating a near miss list of possible word entries for the utterance, each word entry including an associated value of probability that the utterance corresponds to the word entry;

comparing the near-miss list with a stored near-miss confidence template in a collection, wherein each near-miss confidence template in the collection comprises a list of word entries and each word entry in each list includes an associated value; and

performing confidence measure in a speech recognition system based on the comparison of the values in the near-miss list of possible word entries with the values of at least one of the near-miss confidence templates in the collection.

2. The method of

claim 1

wherein the near-miss list of possible word entries comprises a selected number of word entries.

3. The method of

claim 2

wherein each near-miss confidence template in the collection comprises a selected number of word entries, and wherein the near-miss list of possible word entries and each near-miss confidence template in the collection includes a value comprising a compilation of values of a set of word entries.

4. The method of

claim 3

wherein the set of word entries corresponds to word entries not forming part of the selected number of word entries.

5. The method of

claim 4

wherein the step of comparing includes comparing values of like word entries in the near-miss list of possible word entries and the near-miss confidence template, and comparing the values of other word entries in the near-miss list of possible word entries with the value associated with the set of word entries in the near-miss confidence template.

6. The method of

claim 4

wherein the step of comparing includes comparing values of like word entries in the near-miss list of possible word entries and the near-miss confidence template, and comparing the values of other word entries in the near-miss confidence template with the value associated with the set of word entries in the near-miss list of possible word entries.

7. The method of

claim 4

wherein the word entries corresponding to single word entries in the near-miss list of possible word entries and each of the near-miss confidence templates are ordered as a function of numeric value, and the step of comparing includes comparing values of like word entries in the near-miss list of possible word entries and the near-miss confidence template, and comparing the values of other words entries in the near-miss confidence template, having a numeric value greater than the lowest like word entry, with the value associated with the set of word entries in the near-miss list of possible word entries.

8. The method of

claim 1

wherein the near-miss list of possible word entries comprises a number of word entries based on a function of at least one of the values in the near-miss list.

9. The method of

claim 8

wherein the function is a percentage of the value indicating highest probability.

10. The method of

claim 1

and further comprising normalizing the utterance of input speech by the utterance length.

11. The method of

claim 10

and further comprising normalizing the near-miss list of possible word entries.

12. The method of

claim 11

wherein the step of normalizing the near-miss list of possible word entries includes normalizing the values of a selected number of the near-miss list of word entries to sum to a selected total value.

13. The method of

claim 12

wherein the step of normalizing the near-miss list of possible word entries includes normalizing the values of a complete N-Best list of word entries to sum to the selected total value.

14. The method of

claim 11

wherein the step of normalizing the near-miss list of possible word entries includes normalizing the values of the near-miss list as a function of the value of the word entry with the highest probability.

15. The method of

claim 11

wherein the step of normalizing the near-miss list of possible word entries includes normalizing the values of the near-miss list as a function of one of the values in the near-miss list.

16. The method of

claim 15

wherein the step of receiving an utterance includes receiving an utterance of a known word item, and wherein the near-miss list of possible word entries includes an entry and value for the known word item, and wherein step of normalizing includes normalizing the values of the near-miss list of possible word entries as a function of the value for the known word item.

17. The method of

claim 1

wherein the step of comparing includes calculating a confidence measure value, and wherein the step of choosing includes comparing the confidence measure value to a threshold.

18. The method of

claim 17

wherein the step of comparing includes calculating a first confidence measure value when the near-miss list of possible speech entries is compared with a first near-miss confidence template, and calculating a second confidence measure value when the near-miss list of possible word entries is compared with a second near-miss confidence template, and wherein the step of choosing includes comparing a function of the first confidence measure value and the second confidence measure value with a second threshold.

19. The method of

claim 17

wherein the step of comparing includes using a gaussian calculation.

20. A method for generating near-miss confidence templates for a collection in a speech recognition system, the method comprising the steps of:

obtaining a set of near-miss confidence templates for input word tokens of the same word, wherein each near-miss confidence template of the set of near-miss confidence templates comprises a list of word entries having an associated value related to acoustic similarity; and

combining the set of near-miss confidence templates as a function of the lists of word entries and associated values to generate a single near miss confidence template.

21. The method of

claim 20

wherein at least some of the input word tokens correspond to user utterances.

22. The method of

claim 21

wherein the step of obtaining includes normalizing acoustic data related to each utterance by the utterance length.

23. The method of

claim 21

wherein the step of obtaining includes receiving user utterances through a microphone.

24. The method of

claim 20

wherein at least some of the input word tokens correspond to synthetic speech.

25. The method of

claim 24

wherein the step of obtaining includes generating an input word token from stored acoustic data.

26. The method of

claim 25

wherein the step of obtaining includes concatenating stored representative acoustic data to form the input word token.

27. The method of

claim 26

wherein the step of obtaining includes receiving a textual representation of the input word token.

28. The method of

claim 26

wherein the step of concatenating includes referencing a set of decision trees to select representative acoustic data to concatenate.

29. The method of

claim 20

wherein the step of combining includes combining based on a gaussian classifier.

30. The method of

claim 20

wherein the step of obtaining includes normalizing each word item.