Speech processing apparatus and method
AssigneeCanon Kabushiki Kaisha
InventorDavid Llewellyn Rees
Life of this patent
Estimating the term…
Abstract
A speech recognition method and apparatus is disclosed in which outputs a confidence score indicative of the posterior probability of an utterance being correctly matched to a word model. The confidence score for the matching of an utterance to a word model is determined directly from the generated values indicative of the goodness of match between the utterance and stored word models utilizing the following equation: 1 confidence = exp ( - 2 σ S ( x | w ) ) ∑ words exp ( 2 σ S ( x | w ) ) where S(x|w) is the match score for the correlation between a signal x and word w and σ is an experimentally determined constant.
Claims19
A speech recognition apparatus operable to determine a confidence score indicative of the probability that an utterance has been correctly matched to a word, said apparatus comprising:
- a receiver operable to receive a signal indicative of an utterance;
- a matching unit operable to determine, for each of a plurality of word models, a match value indicative of the closeness of match of each said word model to a said signal received by said receiver;
- a calculation unit operable to calculate an exponential of a constant multiplied by a match value determined by said matching unit for an estimate of the likelihood of the receiver receiving said matched signal given that said signal is indicative of an utterance corresponding to the word represented by said word model;
- and
- a processing unit operable to process estimates generated by said calculation unit to determine a confidence score indicative of the probability that an utterance has been correctly matched to a word.
A method of determining a confidence score indicative of the probability that an utterance has been correctly matched to a word, said method comprising the steps of:
- receiving a signal indicative of an utterance;
- determining, for each of a plurality of word models, a match value indicative of the closeness of match of each said word model to said received signal;
- calculating an exponential of a constant multiplied by each said match value as estimates of the likelihood of receiving said matched signal given that said signal is indicative of an utterance corresponding to the word represented by each said word model;
- and
- processing said estimates to determine a confidence score indicative of the probability that an utterance has been correctly matched to a word.
Dependencies are inferred from the claim text.