Speech recognition incorporating a priori probability weighting factors

Claim 3

A method as in claim 1 further comprising the step of: excluding from further repetitive comparison any sequence for which the weighted accumulated measure is, to a degree defined by a pruning criterion, less indicative of similarity than the measures for other such sequences.

Claim 4

A method as in claim 2 wherein the excluding step includes a sub-step of repeatedly adjusting the pruning criterion in dependence upon the number of measures generated and not excluded from further repetitive comparison, such as to tend to maintain that number constant.

Claim 5

Speech recognition apparatus comprising: storage means for storing data relating to reference models representing utterances and data defining allowable sequences of reference models; comparing means to repetitively compare portions of an unknown utterance with reference models to generate, for each of a plurality of allowable sequences of reference models defined by stored data defining such sequences, accumulated measures of similarity including contributions from previously generated measures obtained from comparison of one or more earlier portions of the utterance with a reference model or models in the respective allowable sequence; and weighting means operable to weight the accumulated measures in accordance with predetermined weighting factors representing an a priori probability for each of the allowable sequences wherein the weighting means is operable to weight a measure or accumulated measure for a partial sequence by combined values of the weighting factors for each of the allowable sequences which commences with that partial sequence, modified by any such combined values of the weighting factors applied to a measure generated in respect of a shorter sequence with which that partial sequence commences.

Claim 6

Speech recognition apparatus comprising: storage means for storing data relating to reference models representing utterances and data defining allowable sequences of reference models; comparing means to repetitively compare portions of an unknown utterance with reference models to generate, for each of a plurality of allowable sequences of reference models defined by stored data defining such sequences, accumulated measures of similarity including contributions from previously generated measures obtained from comparison of one or more earlier portions of the utterance with a reference model or models in the respective allowable sequence; and weighting means operable to weight the accumulated measures in accordance with predetermined weighting factors representing an a priori probability for each of the allowable sequences wherein the weighting means is operable to weight a measure or accumulated measure for a partial sequence by combined values of the weighting factors for each of the allowable sequences which commences with that partial sequence, modified by any such combined values of the weighting factors applied to a measure generated in respect of a shorter sequence with which that partial sequence commences.

Claim 7

Apparatus as in claim 4 further comprising: excluding means to exclude from further repetitive comparison any sequence for which the weighted accumulated measure is, to a degree defined by a predetermined pruning criterion, less indicative of similarity than the measures for other such sequences.

Claim 8

Apparatus as in claim 5 wherein the excluding means is arranged to repeatedly adjust the pruning criterion in dependence upon the number of measures generated and not excluded from further repetitive comparison, such as to tend to maintain that number constant.

Claim 11

A method of speech recognition in which a network incorporating weighting factors formed by the method according to claim 7 is used.

Claim 12

A method as in claim 7 comprising the further steps of: assigning the log of the predetermined weighting factors to the final nodes of the network corresponding to the allowable sequences; assigning to each preceding node a log probability value which is the maximum of those values assigned to the node or nodes which follow it; and subtracting from the value for each node the value assigned to the node which precedes it.

Claim 13

A method as in claim 7 wherein the nodes are associated with models representing reference utterances and including the sub-step of modifying parameters of the associated models to reflect the weighting factor assigned to each node.

Claim 14

A method as in claim 7, wherein the recognition network has a tree-structure, at least one node other than the first having more than one branch.

Claim 15

Speech recognition apparatus including a data store containing data relating to a network incorporating weighting factors defined by the method of claim 7.