FALSE ALARM REDUCTION IN SPEECH RECOGNITION SYSTEMS USING CONTEXTUAL INFORMATION

• US 20140207457A1
• Filed: 01/22/2013
• Published: 07/24/2014
• Est. Priority Date: 01/22/2013
• Status: Active Grant

• Current Assignee: Genesys Telecommunications Laboratories Incorporated (Genesys Cloud Services Incorporated)
• Original Assignee: Interactive Intelligence Incorporated (Genesys Cloud Services Incorporated)
• Inventors: Ganapathiraju, Aravind, Wyss, Felix Immanuel, Biatov, Konstantin
• Granted Patent: US 9,646,605 B2
• Time in Patent Office: Days
• US Class Current: 704/243
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 2015/088 Word spotting

Abstract

A system and method are presented for using spoken word verification to reduce false alarms by exploiting global and local contexts on a lexical level, a phoneme level, and on an acoustical level. The reduction of false alarms may occur through a process that determines whether a word has been detected or if it is a false alarm. Training examples are used to generate models of internal and external contexts which are compared to test word examples. The word may be accepted or rejected based on comparison results. Comparison may be performed either at the end of the process or at multiple steps of the process to determine whether the word is rejected.

Claims

1. A computer-implemented method of reducing false alarms in a speech recognition system, the method comprising the steps of:

a. analyzing a context of a set of words;

b. obtaining contexts for said words;

c. obtaining a set of training examples for said words;

d. generating a set of models of the contexts;

e. receiving a set of test words;

f. comparing said set of test words with said set of models;

g. obtaining a threshold for model comparison;

h. determining if a result of comparing said set of test words with a first one of the set of models is within the threshold; and

i. rejecting a word if the result is not within the threshold.

2. The method of claim 1, wherein step (b) comprises defining a size of a local external context and a size of a local internal context of said words.

3. The method of claim 1, wherein step (a) comprises the additional step of defining words from one of:

a grammar specification, a vocabulary, and a set of user defined keywords.

4. The method of claim 1, wherein the method is performed at an acoustical level.

5. The method of claim 1, wherein the method is performed at a phonetic level.

6. The method of claim 1, wherein the comparing of step (f) comprises the additional step of evaluating with a perplexity test a matching test word and a matching model.

7. The method of claim 1, wherein the method is performed at a lexical level.

8. The method of claim 1, wherein step (d) further comprises using sparse vectors to generate said models.

9. The method of claim 1, further comprising the step of evaluating a word match using a distance metric.

10. The distance metric of claim 9, wherein said distance metric comprises cosine distance.

11. A computer-implemented method of reducing false alarms in a speech recognition system, the method comprising the steps of:

a. analyzing a context of a set of words;

b. obtaining contexts for said words;

c. obtaining a set of training examples for said words;

d. generating a set of models of the contexts;

e. receiving a set of test words;

f. comparing said set of test words with a first one of the set of models;

g. obtaining a threshold for model comparison;

h. determining if a result of comparing said set of test words with a first one of the set of models is within the threshold for said first model;

i. rejecting a word if the result is not within the threshold;

j. determining if a result of comparing said set of test words with a second one of the set of models is within a threshold for said second model; and

k. rejecting a word if the result does not meet the second model threshold.

12. The method of claim 11 wherein step (b) comprises the additional step of defining a size of a local external context and a size of a local internal context of said words.

13. The method of claim 11, wherein the method is performed at an acoustical level.

14. The method of claim 11, wherein the method is performed at a phonetic level.

15. The method of claim 11, wherein the comparing of step (f) comprises the additional step of evaluating the matches with a perplexity test.

16. The method of claim 11, wherein step (a) comprises the additional step of analyzing a global context of a set of words.

17. The method of claim 16, wherein the method is performed at a lexical level.

18. The method of claim 16, wherein step (d) further comprises using sparse vectors to generate said models.

19. The method of claim 16 further comprising the step of evaluating a word match using a distance metric.

20. The distance metric of claim 19, wherein said distance metric comprises cosine distance.

21. A computer-implemented method of reducing false alarms in a speech recognition system, the method comprising the steps of:

a. receiving training examples;

b. generating models of acoustical contexts of the training examples;

c. generating models of phonetic contexts of the training examples; and

d. generating models of linguistic contexts of the training examples.

22. The method of claim 21, further comprising the steps of:

a. receiving a set of test examples of words;

b. comparing said test examples of words with said models of acoustical contexts;

c. determining if a result of comparing said test examples of words with said models of acoustical contexts is within a first threshold;

d. comparing said test examples of words with said models of phonetic contexts;

e. determining if a result of comparing said test examples of words with said models of phonetic contexts is within a second threshold;

f. comparing said test examples of words with said models of linguistic contexts; and

g. determining if a result of comparing said test examples of words with said models of linguistic contexts is within a third threshold.

23. The method of claim 21 further comprising the steps of:

a. receiving a set of test examples of words;

b. comparing said test examples of words with at least one of said models of acoustical contexts, phonetic contexts, and linguistic contexts; and

c. determining if a result of comparing said test examples of words with said models of contexts is within a threshold.

24. The method of claim 22 further comprising the additional step of accepting a word if a result of comparing meets a threshold for steps (c), (e), and (g).

25. The method of claim 22 further comprising the additional step of rejecting a word if a result of comparing does not meet a threshold for steps (c), (e), and (g).

26. A system for reducing false alarms in a speech recognition system comprising:

a. means for analyzing a context of a set of words;

b. means for obtaining contexts for said words;

c. means for obtaining a set of training examples for said words;

d. means for generating a set of models of the contexts;

e. means for receiving a set of test words;

f. means for comparing said set of test words with said set of models;

g. means for obtaining a threshold for model comparison;

h. means for determining if a result of comparing said set of test words with a first one of the set of models is within the threshold; and

i. means for rejecting a word if the result is not within the threshold.

27. The system of claim 26 wherein the system additionally comprises:

a. means for testing the word against subsequent models; and

b. means for deciding to accept or reject the word after each test.

28. The system of claim 26 wherein the means for generating a set of models comprises a statistical language model which defines a vocabulary from a word lexicon.

29. The system of claim 28 wherein the statistical language model may comprise one of:

a word grammar, a set of keywords, and a statistical n-gram language model.