EP0540328A2
Expired / Lapsed
EP
⏱️ 0.0 years remaining
Voice recognition
Assignee: International Business Machines Corp
•
Filing: 1992-10-29
•
Granted: —
•
Expires: 2012-10-29
Abstract
Abstract This invention is intended to enable initial ends to be opened irrespective of the configurations of Markov models and without entailing a large increase in amount of processing in Markov model aided voice recognition. For word models dependent on input length, matching paths are started in sequence within the initial end open ranges of sequences of input labels, and the initial likelihood values of the matching paths to be started are set at the greatest median likelihood value of other paths at the start. Matching is carried out in synchronization with input labels, and paths are disbranched in label units by applying beam search techniques to reduce amount of computation. For word models dependent on word model length, matching paths are started in sequence in each state within the initial end open range of each word model, and the initial likelihood values of the matching paths to be started are set at the greatest median likelihood value of other paths at the start. Matching is carried out in synchronization with states, and paths are disbranched in state units by applying beam searching techniques to reduce amount of computation.
Inventors
- 👤 Masafumi Nishimura
Bibliographic Details
Application Number
—
Jurisdiction
EP
Data Source
google_patents
Last Synchronized
2026-11-08 20:40
CPC Classifications
G; G10; G10L; G10L15/00; G10L15/08; G10L15/14; G10L15/142; G09; G09B; G09B19/00; G09B19/04
Claims Analysis
Total Claims: 15
Claim 1
Independent Claim
252 words
1 A method of voice recognition comprising: retaining information on a set of first labels representing an acoustic type assignable to one infinitesimal interval of time, information on the voice model of a vocabulary word expressed as a chained set of Markov models corresponding respectively to a set of second labels representing an acoustic type assignable to another infinitesimal interval of time and information on the probability values of said Markov models; generating corresponding first labels from an input voice sequentially at said infinitesimal intervals of time; matching said voice model of said vocabulary word with said sequence of first labels in synchronization with said infinitesimal intervals of time; accumulating the probability values of said Markov models; and thus obtaining the likelihood that said vocabulary word corresponds to said input voice, characterised by the steps of: setting about matching the word model of a vocabulary word with a sequence of first labels beginning with the first label generated in the applicable interval of time during each of said infinitesimal intervals of time within an initial end open range of said input voice, selecting a predetermined median likelihood value during each of said infinitesimal intervals of time within said initial end open range, setting the initial value of likelihood at said predetermined median likelihood value corresponding to said infinitesimal interval of time when the matching of the voice model of said vocabulary word is started, and abandoning improbable matching paths against said median likelihood value during each of said infinitesimal intervals of time.
Claim 2
Dependent Claim
252 words
1 A method of voice recognition comprising: retaining information on a set of first labels representing an acoustic type assignable to one infinitesimal interval of time, information on the voice model of a vocabulary word expressed as a chained set of Markov models corresponding respectively to a set of second labels representing an acoustic type assignable to another infinitesimal interval of time and information on the probability values of said Markov models; generating corresponding first labels from an input voice sequentially at said infinitesimal intervals of time; matching said voice model of said vocabulary word with said sequence of first labels in synchronization with said infinitesimal intervals of time; accumulating the probability values of said Markov models; and thus obtaining the likelihood that said vocabulary word corresponds to said input voice, characterised by the steps of: setting about matching the word model of a vocabulary word with a sequence of first labels beginning with the first label generated in the applicable interval of time during each of said infinitesimal intervals of time within an initial end open range of said input voice, selecting a predetermined median likelihood value during each of said infinitesimal intervals of time within said initial end open range, setting the initial value of likelihood at said predetermined median likelihood value corresponding to said infinitesimal interval of time when the matching of the voice model of said vocabulary word is started, and abandoning improbable matching paths against said median likelihood value during each of said infinitesimal intervals of time.
Claim 3
Dependent Claim
32 words
A method of voice recognition as claimed in Claim 1 wherein said initial end open range is determined based on detection of one or more initial end candidates of an input voice.
Claim 4
Dependent Claim
66 words
A method of voice recognition as claimed in Claim 2 wherein said initial end open range is determined based on the first initial end candidate, and matching is started with said predetermined median likelihood value at the point of time set as an initial value determined based on the subsequent initial end candidates when a plurality of initial end candidate of an input voice are detected.
Claim 5
Dependent Claim
39 words
A method of voice recognition as claimed in Claim 2 wherein said initial end open range is determined based on each of said initial end candidates when a plurality of initial end candidates of an input voice are detected.
Claim 6
Independent Claim
40 words
A method of voice recognition as claimed in any preceding Claim wherein the matching of the voice model of said vocabulary word is started with each of the Markov models within the initial end open range of the voice model.
Claim 7
Independent Claim
25 words
A method of voice recognition as set forth in any preceding Claim wherein said predetermined median likelihood value is set as the greatest likelihood value.
Claim 8
Independent Claim
268 words
A method of voice recognition comprising retaining information on a set of first labels representing an acoustic type assignable to one infinitesimal interval of time, information on the voice model of a vocabulary word expressed as a Markov model, and information on the probability value of said Markov model; generating corresponding said labels from an input voice sequentially at said infinitesimal intervals of time; matching said voice model of said vocabulary word with said sequence of first labels in synchronization with said infinitesimal intervals of time; accumulating the probability values of said Markov model; and obtaining the likelihood that said vocabulary word corresponds to said input voice, characterised by the steps of: detecting one or more initial ends of said input voice; setting about matching the word model of a vocabulary word with a sequence of first labels beginning with the first label generated in the applicable interval of time during said infinitesimal interval of time corresponding to said one initial end or each of said infinitesimal intervals of time corresponding to said more than one initial end; selecting a predetermined median likelihood value during each of said infinitesimal time intervals corresponding to the second and subsequent initial ends when more than one such initial end occurs; setting the initial likelihood value at said specified median likelihood value corresponding to said infinitesimal interval of time when word model matching is started during each of said infinitesimal intervals of time corresponding to said second and subsequent initial ends; and abandoning improbable matching paths on the basis of said median likelihood value during each of said infinitesimal intervals of time.
Claim 9
Independent Claim
268 words
A method of voice recognition comprising retaining information on a set of first labels representing an acoustic type assignable to one infinitesimal interval of time, information on the voice model of a vocabulary word expressed as a Markov model, and information on the probability value of said Markov model; generating corresponding said labels from an input voice sequentially at said infinitesimal intervals of time; matching said voice model of said vocabulary word with said sequence of first labels in synchronization with said infinitesimal intervals of time; accumulating the probability values of said Markov model; and obtaining the likelihood that said vocabulary word corresponds to said input voice, characterised by the steps of: detecting one or more initial ends of said input voice; setting about matching the word model of a vocabulary word with a sequence of first labels beginning with the first label generated in the applicable interval of time during said infinitesimal interval of time corresponding to said one initial end or each of said infinitesimal intervals of time corresponding to said more than one initial end; selecting a predetermined median likelihood value during each of said infinitesimal time intervals corresponding to the second and subsequent initial ends when more than one such initial end occurs; setting the initial likelihood value at said specified median likelihood value corresponding to said infinitesimal interval of time when word model matching is started during each of said infinitesimal intervals of time corresponding to said second and subsequent initial ends; and abandoning improbable matching paths on the basis of said median likelihood value during each of said infinitesimal intervals of time.
Claim 10
Independent Claim
228 words
A method of voice recognition comprising: retaining information on a set of first labels representing an acoustic type assignable to one infinitesimal interval of time, information on the voice model of a vocabulary word expressed as a Markov model, and information on the probability value of said Markov model; generating corresponding said labels from an input voice sequentially at said infinitesimal intervals of time; matching said voice model of said vocabulary word with said sequence of first labels in synchronization with said infinitesimal intervals of time; accumulating the probability values of said Markov model; and obtaining the likelihood that said vocabulary word corresponds to said input voice, characterised by the steps of: setting about matching the word model of a vocabulary word with a sequence of first labels beginning with the first label generated in the applicable interval of time during each of said infinitesimal intervals of time within the initial end open range of said input voice, selecting a predetermined median likelihood value during each of said infinitesimal intervals of time within said initial end open range, setting the initial likelihood value at said predetermined median likelihood value corresponding to said matching starting infinitesimal interval of time when the matching of said word model is started, and abandoning improbable matching paths on the basis of said median likelihood value during each of said infinitesimal intervals of time.
Claim 11
Independent Claim
228 words
A method of voice recognition comprising: retaining information on a set of first labels representing an acoustic type assignable to one infinitesimal interval of time, information on the voice model of a vocabulary word expressed as a Markov model, and information on the probability value of said Markov model; generating corresponding said labels from an input voice sequentially at said infinitesimal intervals of time; matching said voice model of said vocabulary word with said sequence of first labels in synchronization with said infinitesimal intervals of time; accumulating the probability values of said Markov model; and obtaining the likelihood that said vocabulary word corresponds to said input voice, characterised by the steps of: setting about matching the word model of a vocabulary word with a sequence of first labels beginning with the first label generated in the applicable interval of time during each of said infinitesimal intervals of time within the initial end open range of said input voice, selecting a predetermined median likelihood value during each of said infinitesimal intervals of time within said initial end open range, setting the initial likelihood value at said predetermined median likelihood value corresponding to said matching starting infinitesimal interval of time when the matching of said word model is started, and abandoning improbable matching paths on the basis of said median likelihood value during each of said infinitesimal intervals of time.
Claim 12
Independent Claim
204 words
A method of voice recognition comprising: retaining information on a set of first labels representing an acoustic type assignable to one infinitesimal interval of time, information on the voice model of a vocabulary word expressed as a Markov model, and information on the probability value of said Markov model; generating corresponding said labels from an input voice sequentially at said infinitesimal intervals of time; matching said voice model of said vocabulary word with said sequence of first labels in synchronization with said infinitesimal intervals of time; accumulating the probability values of said Markov model; and obtaining the likelihood that said vocabulary word corresponds to said input voice, characterised by the steps of: setting about matching said sequence of first labels with the remaining part of the applicable voice model in each state within the initial end open range of said voice model, selecting a predetermined median likelihood value in each of said states within said initial end open range, setting the initial likelihood value at said predetermined median likelihood value in said matching starting state when the matching of said sequence of first labels is started, and abandoning improbable matching paths on the basis of said median likelihood value in each of said states.
Claim 13
Independent Claim
204 words
A method of voice recognition comprising: retaining information on a set of first labels representing an acoustic type assignable to one infinitesimal interval of time, information on the voice model of a vocabulary word expressed as a Markov model, and information on the probability value of said Markov model; generating corresponding said labels from an input voice sequentially at said infinitesimal intervals of time; matching said voice model of said vocabulary word with said sequence of first labels in synchronization with said infinitesimal intervals of time; accumulating the probability values of said Markov model; and obtaining the likelihood that said vocabulary word corresponds to said input voice, characterised by the steps of: setting about matching said sequence of first labels with the remaining part of the applicable voice model in each state within the initial end open range of said voice model, selecting a predetermined median likelihood value in each of said states within said initial end open range, setting the initial likelihood value at said predetermined median likelihood value in said matching starting state when the matching of said sequence of first labels is started, and abandoning improbable matching paths on the basis of said median likelihood value in each of said states.
Claim 14
Dependent Claim
268 words
Voice recognition apparatus comprising: means for storing information on a set of first labels representing an acoustic type assignable to one infinitesimal interval of time, information on the voice model of a vocabulary word expressed as a chained set of Markov models corresponding respectively to a set of second labels representing an acoustic type assignable to another infinitesimal interval of time, and information on the probability values of said Markov models; means for generating corresponding first labels from an in put voice sequentially at said infinitesimal intervals of time; means for matching said voice model of said vocabulary word with said sequence of said first labels in synchronization with said infinitesimal intervals of time; means for accumulating the probability values of said Markov models; and means for obtaining the likelihood that said vocabulary word corresponds to said input voice, characterised by: means for setting about matching the word model of a vocabulary word with a sequence of first labels beginning with the first label generated in the applicable interval of time during each of said infinitesimal intervals of time within said initial end open range of said input voice; means for selecting a predetermined median likelihood value during each of said infinitesimal intervals of time within said initial end open range; means for setting the initial value of likelihood at said predetermined median likelihood value corresponding to said infinitesimal interval of time when the matching of the voice model of said vocabulary word is started; and means for abandoning improvable matching paths on the basis of said median likelihood value during each of said infinitesimal intervals of time.
Claim 15
Dependent Claim
268 words
Voice recognition apparatus comprising: means for storing information on a set of first labels representing an acoustic type assignable to one infinitesimal interval of time, information on the voice model of a vocabulary word expressed as a chained set of Markov models corresponding respectively to a set of second labels representing an acoustic type assignable to another infinitesimal interval of time, and information on the probability values of said Markov models; means for generating corresponding first labels from an in put voice sequentially at said infinitesimal intervals of time; means for matching said voice model of said vocabulary word with said sequence of said first labels in synchronization with said infinitesimal intervals of time; means for accumulating the probability values of said Markov models; and means for obtaining the likelihood that said vocabulary word corresponds to said input voice, characterised by: means for setting about matching the word model of a vocabulary word with a sequence of first labels beginning with the first label generated in the applicable interval of time during each of said infinitesimal intervals of time within said initial end open range of said input voice; means for selecting a predetermined median likelihood value during each of said infinitesimal intervals of time within said initial end open range; means for setting the initial value of likelihood at said predetermined median likelihood value corresponding to said infinitesimal interval of time when the matching of the voice model of said vocabulary word is started; and means for abandoning improvable matching paths on the basis of said median likelihood value during each of said infinitesimal intervals of time.
References Cited (Backward Citations)
No citation records indexed.
USPTO Annuity & Maintenance Windows
Maintenance schedule is calculated for granted US utility patents.