Apparatus and method for recognizing speech

Claim 1

An apparatus for recognizing speech, comprising: a means for frequency analyzing an input speech periodically at each input frame, which is a predetermined time interval, for extracting an input pattern composed of time series information of frequency components in a plurality of channels; a means for spectrally normalizing said components of said input pattern in each of the channels in each frame to generate spectrally normalized data W(i,j); a means for detecting start and end points and a speech sound condition of the input speech based on said spectrally normalized data; and a distance calculating and discriminating means comprising: a reference template memory for storing a plurality of reference templates composed of spectrally normalized data Sn(i,kp) in plural frames and plural channels which are produced by frequency analyzing and spectrally normalizing a plurality of predetermined reference speech samples; an input pattern memory for storing the spectrally normalized data W(i,j) in all of the channels for one frame of the input pattern, output from said spectrally normalizing means; a means for starting counting of a speech frame number j of the input pattern after the start point of the input speech has been detected; a means for updating said speech frame number j every input frame while the input speech is in the speech sound condition; a means for holding said speech frame number j while the input speech is in a silent condition; a means for restarting to update said speech frame number j every input frame when the input speech is again in the speech sound condition before the end point of the input speech is detected; a means for setting plural substantially linear matching paths, each having different slopes, between said input pattern and said reference templates by generating plural path-related frame numbers lp, each of which is related to each of said matching paths, in substantially a linear relationship to said speech frame number j according to the equation ##EQU7## where the denominators C p are plural positive real constants each of which is predetermined with respect to each of said matching paths and [] represents a Gaussian symbol, each time said speech frame number j is updated; a means for generating plural reference template frame numbers kp, each of which is related to each of said matching paths, based on said path-related frame numbers lp and the frame length SL(n) of each reference template each time said speech frame number j is updated, according to the following equation; kp=lp (if lp≦SL(n)) SL(n) (if lp>SL(n)) a means for calculating the distance between the spectrally normalized data W(i,j) of said speech frame number j of said input pattern and the spectrally normalized data Sn(i,kp) of said reference template frame number kp of each reference template in each matching path each time said speech frame number j is updated; a means for calculating a dissimilarity relating to each of the matching paths for each of the reference templates by accumulating said distances each time said speech frame number j is updated; a means for detecting a minimum dissimilarity each time said speech frame number j is updated and storing the number of the reference template corresponding to said minimum dissimilarity; and a means for determining the number of said reference template relating to said minimum dissimilarity stored at the speech frame corresponding to the end point of the input speech as the recognized result of the input speech when said end point is detected.

Claim 2

An apparatus for recognizing speech, comprising: a means for frequency analyzing an input speech periodically at each input frame, which is a predetermined time interval, for extracting an input pattern composed of time series information of frequency components in a plurality of channels; a means for spectrally normalizing said components of said input pattern in each of the channels in each frame to generate spectrally normalized data W(i,j); a means for detecting start and end points and a speech sound condition of the input speech based on said spectrally normalized data; and a distance calculating and discriminating means comprising: a reference template memory for storing a plurality of reference templates composed of spectrally normalized data Sn(i,kp) in plural frames and plural channels which are produced by frequency analyzing and spectrally normalizing a plurality of predetermined reference speech samples; an input pattern memory for storing the spectrally normalized data W(i,j) in all of the channels for one frame of the input pattern, output from said spectrally normalizing means; a means for starting counting of a speech frame number j of the input pattern after the start point of the input speech has been detected; a means for updating said speech frame number j every input frame while the input speech is in the speech sound condition; a means for holding said speech frame number j while the input speech is in a silent condition; a means for restarting to update said speech frame number j every input frame when the input speech is again in the speech sound condition before the end point of the input speech is detected; a means for setting plural substantially linear matching paths, each having different slopes, between said input pattern and said reference templates by generating plural path-related frame numbers lp, each of which is related to each of said matching paths, in substantially a linear relationship to said speech frame number j according to the equation ##EQU7## where the denominators C p are plural positive real constants each of which is predetermined with respect to each of said matching paths and [] represents a Gaussian symbol, each time said speech frame number j is updated; a means for generating plural reference template frame numbers kp, each of which is related to each of said matching paths, based on said path-related frame numbers lp and the frame length SL(n) of each reference template each time said speech frame number j is updated, according to the following equation; kp=lp (if lp≦SL(n)) SL(n) (if lp>SL(n)) a means for calculating the distance between the spectrally normalized data W(i,j) of said speech frame number j of said input pattern and the spectrally normalized data Sn(i,kp) of said reference template frame number kp of each reference template in each matching path each time said speech frame number j is updated; a means for calculating a dissimilarity relating to each of the matching paths for each of the reference templates by accumulating said distances each time said speech frame number j is updated; a means for detecting a minimum dissimilarity each time said speech frame number j is updated and storing the number of the reference template corresponding to said minimum dissimilarity; and a means for determining the number of said reference template relating to said minimum dissimilarity stored at the speech frame corresponding to the end point of the input speech as the recognized result of the input speech when said end point is detected.

Claim 9

A method of recognizing speech, comprising the steps of: providing a plurality of reference templates composed of spectrally normalized data in a plurality of frames and a plurality of channels; analyzing the frequencies of an input speech periodically at each input frame, which is a predetermined time interval, to extract an input pattern composed of time series information of frequency components in a plurality of channels; spectrally normalizing said components of said input pattern in each of the channels in each frame to generate spectrally normalized data; starting counting of a speech frame number j of the input pattern after a start point of the input speech is detected; updating said speech frame number j every input frame while the input speech is in the speech sound condition; holding said speech frame number j while the input speech is in a silent condition; restarting to update said speech frame number j every input frame when the input speech is again in the speech sound condition before the end point of the input speech is detected; setting plural substantially linear matching paths, cash having different slopes, between said input pattern and said reference templates by generating plural path-related frame numbers lp, each of which is related to each of said matching paths, in substantially a linear relationship to said speech frame number j according to the equation ##EQU8## where the denominators C p are plural positive real constants each of which is predetermined with respect to each of said matching paths and where [] represents a Gaussian symbol, each time said speech frame number j is updated; generating plural reference template frame numbers kp, each of which is related to each of said matching paths, based on said path-related frame number lp and the frame length SL(n) of each reference template each time said speech frame number j is updated, according to the following equation; kp=lp (if lp≦SL(n)) SL(n) (if lp>(SL(n)) calculating the distance between the spectrally normalized data W(i,j) of said speech frame number j of said input pattern and the spectrally normalized data Sn(i,kp) of said reference template frame number kp of each reference template in each matching path each time said speech frame number j is updated; calculating a dissimilarity relating to each of the matching paths for each of the reference templates by accumulating said distances each time said speech frame number j is updated; detecting a minimum dissimilarity each time said speech frame number j is updated and storing the number of the reference template corresponding to said minimum dissimilarity; and determining the number of said reference template relating to said minimum dissimilarity stored correspondingly to the speech frame number at the end point of the input speech as the recognized result of the input speech when said end point is detected.

Claim 10

A method of recognizing speech, comprising the steps of: providing a plurality of reference templates composed of spectrally normalized data in a plurality of frames and a plurality of channels; analyzing the frequencies of an input speech periodically at each input frame, which is a predetermined time interval, to extract an input pattern composed of time series information of frequency components in a plurality of channels; spectrally normalizing said components of said input pattern in each of the channels in each frame to generate spectrally normalized data; starting counting of a speech frame number j of the input pattern after a start point of the input speech is detected; updating said speech frame number j every input frame while the input speech is in the speech sound condition; holding said speech frame number j while the input speech is in a silent condition; restarting to update said speech frame number j every input frame when the input speech is again in the speech sound condition before the end point of the input speech is detected; setting plural substantially linear matching paths, cash having different slopes, between said input pattern and said reference templates by generating plural path-related frame numbers lp, each of which is related to each of said matching paths, in substantially a linear relationship to said speech frame number j according to the equation ##EQU8## where the denominators C p are plural positive real constants each of which is predetermined with respect to each of said matching paths and where [] represents a Gaussian symbol, each time said speech frame number j is updated; generating plural reference template frame numbers kp, each of which is related to each of said matching paths, based on said path-related frame number lp and the frame length SL(n) of each reference template each time said speech frame number j is updated, according to the following equation; kp=lp (if lp≦SL(n)) SL(n) (if lp>(SL(n)) calculating the distance between the spectrally normalized data W(i,j) of said speech frame number j of said input pattern and the spectrally normalized data Sn(i,kp) of said reference template frame number kp of each reference template in each matching path each time said speech frame number j is updated; calculating a dissimilarity relating to each of the matching paths for each of the reference templates by accumulating said distances each time said speech frame number j is updated; detecting a minimum dissimilarity each time said speech frame number j is updated and storing the number of the reference template corresponding to said minimum dissimilarity; and determining the number of said reference template relating to said minimum dissimilarity stored correspondingly to the speech frame number at the end point of the input speech as the recognized result of the input speech when said end point is detected.