Speech recognition apparatus which predicts word classes from context and words from word classes
First Claim
1. A language generator comprising:
- means for generating two or more word-series hypothesis signals, each word-series hypothesis signal representing a word-series hypothesis comprising a series of linguistic units, each linguistic unit comprising one or more words selected from a language;
means for generating a single-unit hypothesis score signal for each linguistic unit in each word-series hypothesis, each single-unit hypothesis score signal representing a single-unit hypothesis score comprising an estimate of the probability of occurrence of the individual linguistic unit in the word-series hypothesis;
means for generating a word-series hypothesis score signal for each word-series hypothesis, each word-series hypothesis score signal comprising a combination of the single-unit hypothesis scores for the linguistic units in the word-series hypothesis;
means for storing one or more word-series hypothesis signals representing candidate word-series hypotheses having the best word-series hypothesis scores; and
means for outputting at least one word signal representing at least one word in at least one best candidate word-series hypothesis;
characterized in that the means for generating a single-unit hypothesis score signal for a first individual linguistic unit comprises;
means for estimating the conditional probability of occurrence of a first class of linguistic units comprising the first individual linguistic unit, given the occurrence of a context comprising the linguistic units in the word-series hypothesis other than the first individual linguistic unit;
means for estimating the conditional probability of occurrence of the first individual linguistic unit given the occurrence of the first class of linguistic units, and given the occurrence of the context; and
means for generating a single-unit hypothesis score signal for the first individual linguistic unit comprising a first component score comprising a combination of the estimated conditional probability of occurrence of the first class of linguistic units given the occurrence of the context, and the estimated conditional probability of occurrence of the first individual linguistic unit given the occurrence of the first class and given the occurrence of the context.
1 Assignment
0 Petitions
Accused Products
Abstract
A language generator for a speech recognition apparatus scores a word-series hypothesis by combining individual scores for each word in the hypothesis. The hypothesis score for a single word comprises a combination of the estimated conditional probability of occurrence of a first class of words comprising the word being scored, given the occurrence of a context comprising the words in the word-series hypothesis other than the word being scored, and the estimated conditional probability of occurrence of the word being scored given the occurrence of the first class of words, and given the occurrence of the context. An apparatus and method are provided for classifying multiple series of words for the purpose of obtaining useful hypothesis scores in the language generator and speech recognition apparatus.
301 Citations
35 Claims
-
1. A language generator comprising:
-
means for generating two or more word-series hypothesis signals, each word-series hypothesis signal representing a word-series hypothesis comprising a series of linguistic units, each linguistic unit comprising one or more words selected from a language; means for generating a single-unit hypothesis score signal for each linguistic unit in each word-series hypothesis, each single-unit hypothesis score signal representing a single-unit hypothesis score comprising an estimate of the probability of occurrence of the individual linguistic unit in the word-series hypothesis; means for generating a word-series hypothesis score signal for each word-series hypothesis, each word-series hypothesis score signal comprising a combination of the single-unit hypothesis scores for the linguistic units in the word-series hypothesis; means for storing one or more word-series hypothesis signals representing candidate word-series hypotheses having the best word-series hypothesis scores; and means for outputting at least one word signal representing at least one word in at least one best candidate word-series hypothesis; characterized in that the means for generating a single-unit hypothesis score signal for a first individual linguistic unit comprises; means for estimating the conditional probability of occurrence of a first class of linguistic units comprising the first individual linguistic unit, given the occurrence of a context comprising the linguistic units in the word-series hypothesis other than the first individual linguistic unit; means for estimating the conditional probability of occurrence of the first individual linguistic unit given the occurrence of the first class of linguistic units, and given the occurrence of the context; and means for generating a single-unit hypothesis score signal for the first individual linguistic unit comprising a first component score comprising a combination of the estimated conditional probability of occurrence of the first class of linguistic units given the occurrence of the context, and the estimated conditional probability of occurrence of the first individual linguistic unit given the occurrence of the first class and given the occurrence of the context. - View Dependent Claims (2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18)
-
-
19. A method of generating language, said method comprising the steps of:
-
generating two or more word-series hypothesis signals, each word-series hypothesis signal representing a word-series hypothesis comprising a series of linguistic units, each linguistic unit comprising one or more words selected from a language; generating a single-unit hypothesis score signal for each linguistic unit in each word-series hypothesis, each single-unit hypothesis score signal representing a single-unit hypothesis score comprising an estimate of the probability of occurrence of the individual linguistic unit in the word-series hypothesis; generating a word-series hypothesis score signal for each word-series hypothesis, each word-series hypothesis score signal comprising a combination of the single-unit hypothesis scores for the linguistic units in the word-series hypothesis; storing one or more candidate word-series hypothesis signals representing word-series hypotheses having the best word-series hypothesis scores; and outputting at least one word signal representing at least one word in at least one best candidate word-series hypothesis; characterized in that the step of generating a single-unit hypothesis score signal for a first individual linguistic unit comprises; estimating the conditional probability of occurrence of a first class of linguistic units comprising the first individual linguistic unit, given the occurrence of a context comprising the linguistic units in the word-series hypothesis other than the first individual linguistic unit; estimating the conditional probability of occurrence of the first individual linguistic unit given the occurrence of the first class of linguistic units, and given the occurrence of the context; and generating a single-unit hypothesis score signal for the first individual linguistic unit comprising a first component score comprising a combination of the estimated conditional probability of occurrence of the first class of linguistic units given the occurrence of the context, and the estimated conditional probability of occurrence of the first individual linguistic unit given the occurrence of the first class and given the occurrence of the context. - View Dependent Claims (20, 21, 22, 23, 24, 25, 26, 27, 28)
-
-
29. An apparatus for classifying a set of observed events, each event having a predictor feature and a category feature, said predictor feature having one of two or more different possible values, said category feature having one of two or more different possible values, said apparatus comprising:
-
means for measuring the predictor feature value and the category feature value of each event in the set of events; means for selecting one or more different, noncomplementary sets of category feature values, each set containing less than all possible category feature values; means for selecting at least three different, noncomplementary sets of predictor feature values, each set of predictor feature values containing less than all possible predictor feature values; means for generating a primary prediction score for each set of predictor feature values and for each set of category feature values, each primary prediction score comprising an estimate of how well the set of predictor feature values predicts the occurrence of one set of category feature values for all events in the set of observed events; means for classifying an event in a first class if the predictor feature value of the event is a member of the set of predictor feature values having the best primary prediction score; means for generating a secondary prediction score for each set of predictor feature values, other than the set of predictor feature values having the best primary prediction score, and for each set of category feature values, each secondary prediction score comprising an estimate of how well the set of predictor feature values predicts the occurrence of one set of category feature values for all events in the first class of observed events; and means for classifying an event from the first class of events in a first subclass if the predictor feature value of the event is a member of the set of predictor feature values having the best secondary prediction score; characterized in that; each primary prediction score comprises an estimate of how well the set of predictor feature values predicts the occurrence of a set of category feature values in the set of observed events, weighted by a primary weighting factor associated with the set of predictor feature values; each secondary prediction score comprises an estimate of how well the set of predictor feature values predicts the occurrence of a set of category feature values in the first class of observed events, weighted by a secondary weighting factor associated with the set of predictor feature values; and at least one set of predictor feature values, other than the set of predictor feature values having the best primary prediction score, has an associated primary weighting factor which is not equal to a secondary weighting factor associated with the set. - View Dependent Claims (30, 31)
-
-
32. A method of classifying a set of observed events, each event having a predictor feature and a category feature, said predictor feature having one of two or more different possible values, said category feature having one of two or more different possible values, said method comprising the steps of:
-
measuring the predictor feature value and the category feature value of each event in the set of events; selecting one or more different, noncomplementary sets of category feature values, each set containing less than all possible category feature values; selecting at least three different, noncomplementary sets of predictor feature values, each set of predictor feature values containing less than all possible predictor feature values; generating a primary prediction score for each set of predictor feature values and for each set of category feature values, each primary prediction score comprising an estimate of how well the set of predictor feature values predicts the occurrence of one set of category feature values for all events in the set of observed events; classifying an event in a first class if the predictor feature value of the event is a member of the set of predictor feature values having the best primary prediction score; generating a secondary prediction score for each set of predictor feature values, other than the set of predictor feature values having the best primary prediction score, and for each set of category feature values, each secondary prediction score comprising an estimate of how well the set of predictor feature values predicts the occurrence of one set of category feature values for all events in the first class of observed events; and classifying an event from the first class of events in a first subclass if the predictor feature value of the event is a member of the set of predictor feature values having the best secondary prediction score; characterized in that; each primary prediction score comprising an estimate of how well the set of predictor feature values predicts the occurrence of a set of category feature values in the set of observed events, weighted by a primary weighting factor associated with the set of predictor feature values; each secondary prediction score comprises an estimate of how well the set of predictor feature values predicts the occurrence of a set of category feature values in the first class of observed events, weighted by a secondary weighting factor associated with the set of predictor feature values; and at least one set of predictor feature values, other than the set of predictor feature values having the best primary prediction score, has an associated primary weighting factor which is not equal to a secondary weighting factor associated with the set. - View Dependent Claims (33, 34)
-
-
35. A speech recognition apparatus comprising:
-
an acoustic processor for outputting a sequence of coded representations of an utterance to be recognized; means for generating an acoustic match score for each of a plurality of word-series hypotheses, each acoustic match score comprising an estimate of the closeness of a match between a model of the word-series hypothesis and the sequence of coded representations of the utterance, each word-series hypothesis comprising two or more words; means for generating two or more candidate word-series hypothesis signals representing word-series hypotheses having the best acoustic match scores; means for generating a single-word context score for each word in each candidate word-series hypothesis, each single-word context score comprising an estimate of the conditional probability of occurrence of a first individual word given the occurrence of a context set comprising one or more words in the word-series hypothesis other than the first individual word; means for generating a word-series context score for each word-series hypothesis, each word-series context score comprising a combination of the single-word context scores for each word in the word-series hypothesis; means for storing one or more best candidate word-series hypothesis signals representing word-series hypotheses having the best combined acoustic match and context scores; and means for outputting at least one word in at least one best candidate word-series hypothesis; characterized in that the context set is generated by the steps of; providing a set of sequences of two or more words, each word sequence having at least one predictor word and a category word; selecting one or more different, noncomplementary sets of category words, each set of category words containing less than all possible category words; selecting at least three different, noncomplementary sets of predictor words, each set of predictor words containing one or more words in the word-series hypothesis other than the first individual word, each set of predictor words containing less than all possible predictor words; generating a primary prediction score for each set of predictor words and for each set of category words, each primary prediction score comprising an estimate of how well the set of predictor words predicts the occurrence of one set of category words for all word sequences in the set of words sequences; classifying a word sequence in a first class if the predictor word of the word sequence is a member of the set of predictor words having the best primary prediction score; generating a secondary prediction score for each set of predictor words, other than the set of predictor words having the best primary prediction score, and for each set of category words, each secondary prediction score comprising an estimate of how well the set of predictor words predicts the occurrence of one set of category words for all word sequences in the first class of word sequences; and classifying a word sequence from the first class of word sequences in a first subclass if the predictor word of the word sequence is a member of the set of predictor words having the best secondary prediction score; and further characterized in that; each primary prediction score comprises an estimate of how well the set of predictor words predicts the occurrence of a set of category words in the set of word sequences, weighted by a primary weighting factor associated with the set of predictor words; each secondary prediction score comprises an estimate of how well the set of predictor words predicts the occurrence of a set of category words in the first class of word sequences, weighted by a secondary weighting factor associated with the set of predictor words; and at least one set of predictor words, other than the set of predictor words having the best primary prediction score, has an associated primary weighting factor which is not equal to a secondary weighting factor associated with the set.
-
Specification