Method and apparatus for generating a speech-recognition-based call-routing system
View Patent ↗A computerized method is provided for electronically directing a call to a class, such that an utterance spoken by a speaker and received by a call-routing system is classified by the call-routing system as being associated with the class, such that the call-routing system includes a speech-recognition module, a feature-extraction module, and a classification module. The method includes extracting features from recognized speech; weighting elements of a feature vector with respective speech-recognition scores, wherein each weighting element is associated with one of the features; ranking classes to which the features are associated; and electronically directing the call to a highest-ranking class.
1. A computerized method for electronically directing a call to a class, such that an utterance spoken by a speaker and received by a call-routing system is classified by the call-routing system as being associated with the class, such that the call-routing system includes a speech-recognition module, a feature-extraction module, and a classification module, the method comprising:
extracting features from recognized speech;
weighting elements of a feature vector with respective speech-recognition scores, wherein each element is associated with one of the features and each element is a number of times the element's associated feature is included in the recognized speech;
ranking classes to which the features are associated based at least in part on the elements of the feature vector; and
electronically directing the call to a highest-ranking class.
2. The method of claim 1 , wherein ranking classes includes generating a score for each class and ranking the classes in a ranked list of the classes, and wherein the scores determine the ranking of the classes in the ranked list.
3. The method of claim 1 , further comprising generating the recognized speech from the utterance.
4. The method of claim 1 , further comprising generating speech-recognition scores for the features.
5. The method of claim 1 , wherein generating speech-recognition scores for the features includes generating a confidence score for each word of the recognized speech.
6. The method of claim 5 , wherein a confidence score for a feature that includes a phrase is an average of the speech-recognition scores for each word forming the phrase.
7. The method of claim 5 , wherein a confidence score for a feature that includes a word is the confidence score for the word.
8. The method of claim 1 , wherein the confidence scores are in the range of 0 to 1.
9. A computerized method for electronically directing a call to a class, such that an utterance spoken by a speaker and received by a call-routing system is classified by the call-routing system as being associated with the class, such that the call-routing system includes a speech-recognition module, a feature-extraction module, and a classification module, the method comprising:
extracting sets of features from respective speech-recognition results;
generating a feature vector for each set of features, wherein:
each feature vector includes a set of elements,
each element of a feature vector is associated with a corresponding one of the features in the sets of features associated with the feature vector, and
each element is a number of times the element's associated feature is included in an associated speech-recognition result;
weighting elements of each feature vector with respective weighting factors;
summing the feature vectors to generate a final feature vector;
ranking classes associated with the features based on the final feature vector; and
electronically directing the call to a highest-ranking class.
10. The method of claim 9 , further comprising deriving the speech-recognition results from the utterance.
11. The method of claim 10 , wherein the speech-recognition results are generated by the speech-recognition module.
12. The method of claim 9 , wherein the speech-recognition results are a set of nbest speech-recognition results.
13. The method of claim 9 , further comprising generating the weighting factors for the features.
14. The method of claim 9 , further comprising deriving the weighting factors from speech-recognition scores for words in the recognized speech results.
15. The method of claim 14 , wherein each weighting factor is generated from the equation
W
[
i
]
=
exp
[
α
(
l_score
(
i
)
-
max_l
_score
)
]
∑
j
=
1
N
exp
[
α
(
l_score
(
j
)
-
max_l
_score
)
]
,
wherein 1_score(i) is the speech-recognition score of for a given sequence of words for which a weight is generated, max — 1_score is a maximum speech-recognition score of the nbest sequences of words, 1_score(j) is the acoustic confidence for a jth nbest sequence of words, and α is a tunable parameter.
16. The method of claim 9 , wherein an acoustic-confidences score for a feature that includes a word is the speech-recognition score for the word.
17. The method of claim 16 , wherein in the speech-recognition scores are in the range of 0 to 1.
18. A computerized method for electronically directing a call to a class, such that an utterance spoken by a speaker and received by a call-routing system is classified by the call-routing system as being associated with the class, such that the call-routing system includes a speech-recognition module, a feature-extraction module, and a classification module, the method comprising:
extracting sets of features from respective speech-recognition results;
generating a feature vector for each set of features, wherein:
each feature vector includes a set of elements,
each element of a feature vector is associated with a corresponding one of the features in the sets of features associated with the feature vector, and
each element is a number of times the element's associated feature is included in an associated speech-recognition result;
generating a ranked list for each feature vector based at least in part on the elements of each feature vector;
combining the ranked lists; and
electronically directing the call to a highest scoring class.
19. The method of claim 18 , wherein the speech-recognition results are generated by the speech-recognition module based on the utterance.
20. The method of claim 18 , wherein the speech-recognition results include a set of nbest speech-recognition results.
21. The method of claim 18 , further comprising generating a classifier score for each class in each ranked list.
22. The method of claim 21 , further comprising weighting each classifier score with a speech-recognition score.
23. The method of claim 18 , wherein combining the ranked lists includes generating final-classifier scores, and wherein the highest scoring class has a highest-final-classifier score.
24. The method of claim 23 , wherein the final-classifier scores are derived from the formula:
final_classifier score(class — i )=Σ j=1 n classifier_score — j (class — i ) α ·accoustic_confidence_score — j β ,
wherein class_i is a given class for which a final-classifier score is calculated, the sum is over nbest sequences of words, the classifier_score_j(class_i) is a classifier score for a class associated with a jth nbest sequence of words, and the acoustic_confidence_score_j is the speech-recognition score for the jth nbest sequence of words, and α and β are tunable parameter used to place emphasis on the classifier_scores_j or the acoustic_confidence_scores_j.
25. A call-routing system for electronically directing a call to a class, the system comprising:
a speech-recognition module configured to generate nbest sequences of words based on an utterance spoken by a caller;
a feature-extraction module configured to generate features vectors for the nbest sequences of words, wherein:
each feature vector includes a set of elements,
each element of a feature vector is associated with a corresponding one of the features of the feature vector, and
each element is a number of times the element's associated feature is included in an associated speech-recognition result in the nbest sequences of words; and
a classification module configured to form a ranked list of classes for the feature vectors based at least in part on the elements of each feature vector, and direct the call to a highest-ranking class in the ranked list of classes.
26. The system of claim 25 , wherein the classification module is further configured to generate a set of ranked lists for the feature vectors, and wherein the first mentioned ranked list is a summation of the ranked lists of the set of ranked lists.
27. The system of claim 26 , wherein the classification module is configured to weight the ranked lists of the set of ranked lists with weighting factors.
28. The system of claim 27 , wherein the weighting factors are calculated from the equation
W
[
i
]
=
exp
[
α
(
l_score
(
i
)
-
max_l
_score
)
]
∑
j
=
1
N
exp
[
α
(
l_score
(
j
)
-
max_l
_score
)
]
,
wherein 1_score(i) is the speech-recognition score of for a given sequence of words for which a weight is generated, max — 1_score is a maximum speech-recognition score of the nbest sequences of words, 1_score(j) is the acoustic confidence for a jth nbest sequence of words, and α is a tunable parameter.
29. The system of claim 25 , wherein the feature-extraction module is configured to weight elements of the feature vectors with acoustic-confidences scores and sum the feature vectors to form a final feature vector.
30. The system of claim 29 , wherein the classification module is configured to form the ranked list from the final feature vector.