Phoneme labelling of frame by accumulating evidence obtained from spotting results of microscopic units

Claim 2

A speech recognition system comprising: speech input means for converting a speech input into an electrical signal; analysis means for analyzing an input pattern of said electrical signal received from said speech input means; memory means for storing a plurality of standard patterns, each of which represents a macroscopic unit composed of connected phoneme patterns and wherein each phoneme pattern occupies a plurality of spectrum frames and is formed by a plurality of joined spectrum frame sequences, each said phoneme pattern including at least one of a phoneme and a phonemic particle, and each frame of each said phoneme pattern is labelled by a label expressing a category name of one of a phoneme or phonemic particle; continuous dynamic programming calculating means for calculating a distance between said input pattern analyzed by said analysis means and each of said plurality of standard patterns stored in said memory means, and for calculating a plurality of matching values which respectively correspond to each calculated said distance between said input pattern and said plurality of standard patterns based on a matching system which includes continuous dynamic programming; parameter calculating means for calculating a parameter expressing a matching history of labels as a function of time between said input pattern and each one of said plurality of standard patterns which correspond to respective ones of said plurality of matching values selected using said continuous dynamic programming executed by said continuous dynamic programming calculating means; comparison means for comparing each of said plurality of matching values calculated by said continuous dynamic programming calculating means with a predetermined threshold value and for determining when a given one of said plurality of matching values is smaller than said predetermined threshold value; and recognition means for determining which said phoneme pattern, expressing said at least one of a phoneme and a phonemic particle, corresponds to said parameter calculated by said parameter calculating means for said given one of said plurality of matching values, and wherein a label for each input frame is determined by accumulating the number of labels with the same frame time and also with the same name in a plurality of said matching histories of labels.

Claim 3

A speech recognition system as claimed in claim 1, wherein said analysis means includes a plurality of band-pass filters, each of which receives said electrical signal from said speech input means to produce an output analysis for each of a corresponding plurality of predetermined frequency domains.

Claim 4

A speech recognition system as claimed in claim 1, wherein said analysis means further comprises means for performing a Fourier conversion on said electrical signal from said speech input means for each of a plurality of predetermined frequency domains.

Claim 5

A speech recognition system as claimed in claim 1, wherein said memory means contains in storage information representing a plurality of phonemes and phonemic particles which were obtained by said recognition means from continuous speech input.

Claim 6

A speech recognition system comprising: speech input means for converting a speech input into an electrical signal; analysis means for analyzing an input pattern of said electrical signal received from said speech input means; memory means for storing a plurality of standard patterns, each of which represents a macroscopic unit composed of connected phoneme patterns and wherein each phoneme pattern occupies a plurality of spectrum frames and is formed by a plurality of joined spectrum frame sequences, each said phoneme pattern including at least one of a phoneme and a phonemic particle, and each frame of each said phoneme pattern is labelled by a label expressing a category name of a phoneme or phonemic particle; continuous dynamic programming calculating means for calculating a distance between said input pattern analyzed by said analysis means and each of said plurality of standard patterns stored in said memory means, and for calculating a plurality of matching values which respectively correspond to each calculated said distance between said input pattern and said plurality of standard patterns based on a matching system which includes continuous dynamic programming; parameter calculating means for calculating parameters expressing a matching history as a function of time between said input pattern and each one of said plurality of standard patterns which correspond to respective ones of said plurality of matching values selected using said continuous dynamic programming executed by said continuous dynamic programming calculating means; said matching history expressed by said parameter comprising a plurality of frames having a time sequence, and each said frame having a corresponding phoneme label; comparison means for comparing each of said plurality of matching values calculated by said continuous dynamic programming calculating means with a predetermined threshold value and determining when a given one of said plurality of matching values is smaller than said predetermined threshold value; and accumulating means for accumulating, for each frame of a matching history, the number of corresponding ones of said phoneme patterns, by selecting only those ones of said parameters calculated by said parameter calculating means which correspond to the respective ones of said matching values which are determined by said comparison means to be smaller than said threshold value, the selected said parameters corresponding to respective ones of said standard patterns; and recognition means for determining one of said phoneme labels used for said standard patterns for which the number of occurrences accumulated by said accumulating means is maximum as recognition phonemes of said input pattern at said frame and wherein accumulation is performed by calculating said parameters composed both of parameters made at a current frame and parameters made at past frames including labels with the same frame time.

Claim 7

A speech recognition system comprising: speech input means for converting a speech input into an electrical signal; analysis means for analyzing an input pattern of said electrical signal received from said speech input means; memory means for storing a plurality of standard patterns, each of which represents a macroscopic unit composed of connected phoneme patterns and wherein each phoneme pattern occupies a plurality of spectrum frames and is formed by a plurality of joined spectrum frame sequences, each said phoneme pattern including at least one of a phoneme and a phonemic particle, and each frame of each said phoneme pattern is labelled by a label expressing a category name of a phoneme or phonemic particle; continuous dynamic programming calculating means for calculating a distance between said input pattern analyzed by said analysis means and each of said plurality of standard patterns stored in said memory means, and for calculating a plurality of matching values which respectively correspond to each calculated said distance between said input pattern and said plurality of standard patterns based on a matching system which includes continuous dynamic programming; parameter calculating means for calculating parameters expressing a matching history as a function of time between said input pattern and each one of said plurality of standard patterns which correspond to respective ones of said plurality of matching values selected using said continuous dynamic programming executed by said continuous dynamic programming calculating means; said matching history expressed by said parameter comprising a plurality of frames having a time sequence, and each said frame having a corresponding phoneme label; comparison means for comparing each of said plurality of matching values calculated by said continuous dynamic programming calculating means with a predetermined threshold value and determining when a given one of said plurality of matching values is smaller than said predetermined threshold value; and accumulating means for accumulating, for each frame of a matching history, the number of corresponding ones of said phoneme patterns, by selecting only those ones of said parameters calculated by said parameter calculating means which correspond to the respective ones of said matching values which are determined by said comparison means to be smaller than said threshold value, the selected said parameters corresponding to respective ones of said standard patterns; and recognition means for determining one of said phoneme labels used for said standard patterns for which the number of occurrences accumulated by said accumulating means is maximum as recognition phonemes of said input pattern at said frame and wherein accumulation is performed by calculating said parameters composed both of parameters made at a current frame and parameters made at past frames including labels with the same frame time.

Claim 8

A speech recognition system as claimed in claim 5, wherein said analysis means includes a plurality of band-pass filters, each of which receives said electrical signal from said speech input means to produce an output analysis for each of a corresponding plurality of predetermined frequency domains.

Claim 9

A speech recognition system as claimed in claim 5, wherein said analysis means further comprises means for performing a Fourier conversion on said electrical signal from said speech input means for each of a plurality of predetermined frequency domains.

Claim 10

A speech recognition system as claimed in claim 5, wherein said memory means contains in storage information representing a plurality of phonemes and phonemic particles which were obtained by said recognition means from continuous speech input.