Automatic generation of simple Markov model stunted baseforms for words in a vocabulary

Claim 2

In a system that (i) generates in an acoustic processor a string of fenemes in response to speech input, (ii) defines each word in a vocabulary by a respective fenemic baseform comprising a sequence of Markov model fenemic phones each of which relate to a feneme generatable by the acoustic processor, and (iii) defines an alphabet of composite Markov model phonetic phones each of which relates to a phonetic element and each of which correlates to at least one fenemic phone, a method of selecting Markov model phones for inclusion in a stunted baseform constructed of a sequence of phonetic phones, the method comprising the steps of: (a) generating a respective string of fenemes in response to an utterance of a selected word; (b) selecting the fenemic baseform having the highest probability of producing the string of fenemes and aligning the selected fenemic baseform against the generated string of fenemes; (c) replacing each fenemic phone in the fenemic baseform by a composite phone corresponding thereto, thereby forming a composite phone baseform; (d) selecting a pair of adjacent composite phones in the composite phone baseform; (e) determining a substring of fenemes aligned against the selected pair of adjacent phones; (f) computing, for each composite phone in the composite phone alphabet, a respective probability of producing the determined substring of fenemes; (g) selecting the composite phone having the highest probability of having produced the determined substring; (h) repeating steps (d) through (g) for each pair of adjacent phones in the composite phone baseform; and (j) replacing at least one pair of adjacent composite phones by the selected composite phone corresponding thereto.

Claim 3

The method of claim 1 comprising the further step of: (k) measuring, for each pair of adjacent phones, an adverse effect on the probability of producing fenemes in the determined substring in the event that the pair of adjacent composite phones aligned against the determined substring is replaced by the selected composite phone corresponding thereto; wherein step (j) includes the steps of choosing for replacement only those selected composite phones having adverse effect within prescribed limitations.

Claim 4

The method of claim 2 comprising the further step of: (m) defining a new baseform following step (k) and repeating steps (d) through (k) with the new baseform; and (n) repeating step (m) until the defined new baseform is reduced to a predefined number of composite phones, thereby providing a stunted baseform.

Claim 5

The method of claim 3 wherein step (k) includes the step of: (p) replacing only one pair of adjacent phones by the composite phone corresponding thereto, the replacing of said one pair of adjacent phones by the composite phone corresponding thereto yielding the least measured adverse effect of all replacements; the new baseform in step (m) having one less phone than the baseform prior to the replacing step (p).

Claim 6

The method of claim 4 wherein the measuring of adverse effect includes the step of: (q) measuring the reduction in probability between (i) the probability of each of the adjacent phones producing the fenemes aligned thereagainst and (ii) the probability of the single composite phone producing the fenemes aligned against both adjacent phones.

Claim 7

The method of claim 5 wherein step (q) comprises the steps of: (r) determining a probability of the first phone in a given adjacent pair of phones producing the fenemes aligned thereagainst as L m and determining the probability of the second phone in the given pair of adjacent phones producing the fenemes aligned thereagainst as L n ; determining the probability of the composite phone corresponding to the given pair of adjacent phones producing the substring of fenemes aligned against the given pair of adjacent phones as L mn ; and comparing the probabilities L m and L n with the probability L mn .

Claim 8

The method of claim 6 wherein L m , L n , and L mn are logarithmic probabilities and the measured reduction in probability is: L.sub.m +L.sub.n -L.sub.mn

Claim 9

The method of claim 3 wherein the alphabet of composite phones represents a phoneme alphabet and wherein the stunted baseform of step (n) provides a phonetic representation of the word corresponding to the constructed baseform.

Claim 10

In a system which processes speech, a method of automatically constructing for each word in a vocabulary a stunted baseform, the method comprising the steps of: (a) defining each word in the vocabulary as a respective fenemic baseform formed of a sequence of fenemic phones, each fenemic phone being from an alphabet of N fenemic phones, and entering each fenemic baseform into storage; (b) defining an alphabet of composite phones, each composite phone being associated with at least one fenemic phone; (c) generating, in an acoustic processor which associates one feneme in a fixed set of fenemes to each successive interval of speech in an utterance of speech, a plurality of feneme strings for a selected word, each feneme string being generated in response to an utterance of the selected word; (d) selecting the stored fenemic baseform having the highest joint probability of having produced all the generated feneme strings for the selected word and aligning the selected fenemic baseform against each of the feneme strings; (e) replacing each fenemic phone by a composite phone corresponding thereto, thereby forming a composite phone baseform; (f) selecting a pair of adjacent phones in the composite phone baseform; (g) for each feneme string, aligning a substring of fenemes against the selected pair of adjacent phones; (h) selecting a composite phone for a given pair of adjacent composite phones which has the highest joint probability of producing all of the substrings aligned against said given pair of adjacent composite phones; (j) repeating steps (f) through (h) for all pairs of adjacent phones, thereby providing a selected composite phone for each adjacent pair of phones in the composite phone baseform; and (k) determining which selected composite phone yields the least adverse effect on the probability of producing fenemes in the aligned substring therefor in the event that the pair of adjacent phones is replaced by the selected composite phone corresponding thereto.

Claim 11

In a system which processes speech, a method of automatically constructing for each word in a vocabulary a stunted baseform, the method comprising the steps of: (a) defining each word in the vocabulary as a respective fenemic baseform formed of a sequence of fenemic phones, each fenemic phone being from an alphabet of N fenemic phones, and entering each fenemic baseform into storage; (b) defining an alphabet of composite phones, each composite phone being associated with at least one fenemic phone; (c) generating, in an acoustic processor which associates one feneme in a fixed set of fenemes to each successive interval of speech in an utterance of speech, a plurality of feneme strings for a selected word, each feneme string being generated in response to an utterance of the selected word; (d) selecting the stored fenemic baseform having the highest joint probability of having produced all the generated feneme strings for the selected word and aligning the selected fenemic baseform against each of the feneme strings; (e) replacing each fenemic phone by a composite phone corresponding thereto, thereby forming a composite phone baseform; (f) selecting a pair of adjacent phones in the composite phone baseform; (g) for each feneme string, aligning a substring of fenemes against the selected pair of adjacent phones; (h) selecting a composite phone for a given pair of adjacent composite phones which has the highest joint probability of producing all of the substrings aligned against said given pair of adjacent composite phones; (j) repeating steps (f) through (h) for all pairs of adjacent phones, thereby providing a selected composite phone for each adjacent pair of phones in the composite phone baseform; and (k) determining which selected composite phone yields the least adverse effect on the probability of producing fenemes in the aligned substring therefor in the event that the pair of adjacent phones is replaced by the selected composite phone corresponding thereto.

Claim 12

The method of claim 9 comprising the further step of: (m) setting an adverse effect threshold and determining if the set threshold is exceeded if the composite phone yielding the least adverse effect replaces the adjacent phones corresponding thereto.

Claim 13

The method of claim 10 comprising the further step of: (n) replacing the selected composite phone yielding the least adverse effect for the adjacent phones corresponding thereto, thereby forming a new baseform.

Claim 14

The method of claim 11 comprising the further step of: (p) setting a desired length of phones for the stunted baseform; and (r) repeating steps (f) through (n) until either the length of phones in the baseform is reduced to the desired length or until the adverse effect threshold is exceeded by replacing an adjacent pair of phones by the selected composite phone which yields the least adverse effect upon replacement, step (n) being omitted if the threshold in step (m) is exceeded.

Claim 15

The method of claim 12 wherein step (k) includes the step of: (t) performing a paired t-test which, for each substring aligned against the selected pair of adjacent phones, includes the step of comparing (i) the probability of producing a given feneme in an aligned substring by the adjacent phone aligned thereagainst with (ii) the probability of producing the given feneme in the substring by the selected composite phone.

Claim 16

The method of claim 13 comprising the further step of: (u) repeating step (t) for each feneme in each substring aligned against the selected pair of adjacent phones.

Claim 17

The method of claim 12 wherein step (k) includes the steps of: (v) measuring the adverse effect as the value of the probability reduction resulting from replacing an adjacent pair of phones by the selected composite phone corresponding thereto.

Claim 18

The method of claim 12 comprising the further step of: (w) repeating steps (c) through (r) for each word in the vocabulary.

Claim 19

The method of claim 16 in a speech recognition system, the method including the further step of: (x) storing the stunted baseforms in a memory of the speech recognition system for use in performing acoustic matching.

Claim 20

The method of claim 9 wherein step (b) includes the steps of: (aa) selecting a pair of phones from the alphabet of fenemic phones; (bb) determining the entropy for each phone in the selected pair, wherein entropy for each phone is defined by the expression: entropy=-Σp.sub.i log.sub.2 p.sub.i the sum being taken over all fenemes (i=1,N) where N is the number of fenemes in the alphabet and where each p i is the probability of a particular phone producing the feneme i; (cc) determining the entropy of a composite phone resulting from the merger of the two phones of the selected pair into a single phone wherein each output probability associated with the composite phone is the average of the output probabilities of the two phones that are paired; (dd) determining the loss of phone/feneme mutual information based on an entropy measure for the composite phone and each merged phone according to the expression: Loss=E.sub.composite -(E.sub.phone1 +E.sub.phone2)/2 wherein the larger the loss of mutual information, the less the composite phone reflects the two paired phones; (ee) repeating steps (aa) through (dd) for each pair of fenemic phones; and (ff) selecting the composite phone representing the smallest loss for a given selected pair of merged phones; and (gg) repeating steps (aa) through (ff) until the number of composite phones in the composite phone alphabet is n, where n is an integer less than N.

Claim 21

In a speech recognition system including means for performing acoustic matching between (i) stunted word baseforms each comprised of a sequence of Markov model phones and (ii) a string of fenemes generated by an acoustic processor in response to speech input, a method of automatically constructing for each word in a vocabulary a stunted baseform used in the acoustic matching, the method comprising the steps of: (a) defining each word in the vocabulary as a respective fenemic baseform formed of a sequence of fenemic phones, each fenemic phone being from an alphabet of N fenemic phones, and entering each fenemic baseform into storage; (b) defining an alphabet of composite phones, each composite phone being associated with at least one fenemic phone; (c) generating a plurality of feneme strings for a selected word, each feneme string being generated in response to an utterance of the selected word; (d) selecting the stored fenemic baseform having the highest joint probability of producing all the generated feneme strings for the subject word and aligning the fenemic phones of the selected fenemic baseform against each of the feneme strings; (e) replacing each fenemic phone by a composite phone corresponding thereto, thereby forming a composite phone baseform; (f) selecting a pair of adjacent composite phones in the composite phone baseform; (g) for each generated feneme string, aligning a substring of the generated feneme string against the selected pair of adjacent composite phones; (h) selecting the composite phone in the alphabet thereof which has the highest joint probability of producing all of the determined substrings; (j) repeating steps (f) through (h) for all pairs of adjacent composite phones, thereby providing a selected composite phone for each adjacent pair of composite phones in the baseform; (k) determining which single composite phone yields the least adverse effect when replacing the adjacent pair of composite phones corresponding thereto; and (m) replacing the selected composite phone yielding the least adverse effect for the adjacent composite phones corresponding thereto, thereby forming a new baseform.

Claim 22

In a speech recognition system including means for performing acoustic matching between (i) stunted word baseforms each comprised of a sequence of Markov model phones and (ii) a string of fenemes generated by an acoustic processor in response to speech input, a method of automatically constructing for each word in a vocabulary a stunted baseform used in the acoustic matching, the method comprising the steps of: (a) defining each word in the vocabulary as a respective fenemic baseform formed of a sequence of fenemic phones, each fenemic phone being from an alphabet of N fenemic phones, and entering each fenemic baseform into storage; (b) defining an alphabet of composite phones, each composite phone being associated with at least one fenemic phone; (c) generating a plurality of feneme strings for a selected word, each feneme string being generated in response to an utterance of the selected word; (d) selecting the stored fenemic baseform having the highest joint probability of producing all the generated feneme strings for the subject word and aligning the fenemic phones of the selected fenemic baseform against each of the feneme strings; (e) replacing each fenemic phone by a composite phone corresponding thereto, thereby forming a composite phone baseform; (f) selecting a pair of adjacent composite phones in the composite phone baseform; (g) for each generated feneme string, aligning a substring of the generated feneme string against the selected pair of adjacent composite phones; (h) selecting the composite phone in the alphabet thereof which has the highest joint probability of producing all of the determined substrings; (j) repeating steps (f) through (h) for all pairs of adjacent composite phones, thereby providing a selected composite phone for each adjacent pair of composite phones in the baseform; (k) determining which single composite phone yields the least adverse effect when replacing the adjacent pair of composite phones corresponding thereto; and (m) replacing the selected composite phone yielding the least adverse effect for the adjacent composite phones corresponding thereto, thereby forming a new baseform.

Claim 23

The method of claim 19 comprising the further step of: (n) setting a desired length of phones for the stunted baseform; and (p) repeating steps (f) through (m) until either the length of phones in the baseform is reduced to the desired length or until the adverse effect threshold is exceeded by replacing an adjacent pair of phones by the selected composite phone which yields the least adverse effect upon replacement, step (m) being omitted if an adverse effect threshold is exceeded.

Claim 24

The method of claim 20 comprising the further step of: (q) arranging the stunted baseforms for the words in the vocabulary into a tree structure in storage portion of the speech recognition system.

Claim 25

In a system that (i) defines each word in a vocabulary by a fenemic baseform of fenemic phones, (ii) defines an alphabet of composite phones each of which corresponds to at least one fenemic phone, and (iii) generates a string of fenemes in response to speech input, a method of constructing a stunted phonetic-type word baseform of phones comprising the step of: converting a fenemic phone word baseform into a stunted word baseform of composite phones including the steps of: (a) replacing each fenemic phone in the fenemic phone word baseform by the composite phone corresponding thereto; and (b) replacing at least one pair of adjacent composite phones by a single composite phone where the adverse effect of the replacing is below a predefined threshold.

Claim 26

In a system that (i) defines each word in a vocabulary by a fenemic baseform of fenemic phones, (ii) defines an alphabet of composite phones each of which corresponds to at least one fenemic phone, and (iii) generates a string of fenemes in response to speech input, a method of constructing a stunted phonetic-type word baseform of phones comprising the step of: converting a fenemic phone word baseform into a stunted word baseform of composite phones including the steps of: (a) replacing each fenemic phone in the fenemic phone word baseform by the composite phone corresponding thereto; and (b) replacing at least one pair of adjacent composite phones by a single composite phone where the adverse effect of the replacing is below a predefined threshold.

Claim 27

The method of claim 22 wherein step (b) includes the steps of replacing one pair of adjacent composite phones by a single composite phone, wherein said replacing results in less adverse effect than the replacing of any other two adjacent compozsite phones by a single composite phone.

Claim 28

The method of claim 23 comprising the further step of: (c) defining a new baseform following step (b) and repeating steps (a) and (b) until a word baseform of a given length is achieved.