US5822728A — Multistage word recognizer based on reliably dete…

Claim 1

Independent Claim 185 words

A word recognition processor for processing an input speech utterance in a speech recognition system, comprising: a phoneme similarity module receptive of said input speech utterance for producing phone similarity data indicative of the correlation between said input speech utterance and predetermined phone model speech data; a high similarity module coupled to said phoneme similarity module for identifying those regions of the phone similarity data that exceed a predetermined threshold; a region count stage having a first word prototype database for storing similarity region count data for a plurality of predetermined words; said region count stage coupled to said high similarity module and generating a first list of word candidates selected from said first word prototype database based on similarity regions; a target congruence stage having a second word prototype database for storing word prototype data corresponding to a said plurality of predetermined words; said target congruence stage being receptive of said first list of word candidates and being coupled to said high similarity module for generating a second list of at least one word candidate, selected from said first list based on similarity regions.

Claim 2

Dependent Claim 185 words

A word recognition processor for processing an input speech utterance in a speech recognition system, comprising: a phoneme similarity module receptive of said input speech utterance for producing phone similarity data indicative of the correlation between said input speech utterance and predetermined phone model speech data; a high similarity module coupled to said phoneme similarity module for identifying those regions of the phone similarity data that exceed a predetermined threshold; a region count stage having a first word prototype database for storing similarity region count data for a plurality of predetermined words; said region count stage coupled to said high similarity module and generating a first list of word candidates selected from said first word prototype database based on similarity regions; a target congruence stage having a second word prototype database for storing word prototype data corresponding to a said plurality of predetermined words; said target congruence stage being receptive of said first list of word candidates and being coupled to said high similarity module for generating a second list of at least one word candidate, selected from said first list based on similarity regions.

Claim 3

Dependent Claim 47 words

The word recognition processor of claim 1 further comprising a fine match stage having word template database for storing word template data corresponding to said plurality of predetermined words; said fine match stage being receptive of said second list of word candidates for selecting the recognized word.

Claim 4

Dependent Claim 37 words

The word recognition processor of claim 1 wherein said phoneme similarity module includes a phone model database for storing phone model speech data corresponding to a plurality of phonemes that comprise said predetermined phone model speech data.

Claim 5

Dependent Claim 35 words

The word recognition processor of claim 1 wherein said region count stage produces a first score corresponding to the degree of fit between the input utterance and each of the first list of word candidates.

Claim 6

Dependent Claim 35 words

The word recognition processor of claim 1 wherein said target congruence stage produces a second score corresponding to the degree of fit between the input utterance and each of the second list of word candidates.

Claim 7

Dependent Claim 88 words

The word recognition processor of claim 1 wherein said region count stage produces a first score corresponding to the degree of fit between the input utterance and each of the first list of word candidates; wherein said target congruence stage produces a second score corresponding to the degree of fit between the input utterance and each of the second list of word candidates; and wherein said recognizer combines the first and second scores and selects at least the word with the best score as a final word candidate.

Claim 8

Dependent Claim 18 words

The word recognition processor of claim 6 wherein said processor combines the first and second scores by averaging.

Claim 9

Dependent Claim 25 words

The word recognition processor of claim 1 wherein said high similarity module produces a parameterized representation of high similarity regions of the phone similarity data.

Claim 10

Dependent Claim 23 words

The word recognition processor of claim 8 wherein said parameterized representation includes a representation of the phone similarity peak location and peak height.

Claim 11

Dependent Claim 29 words

The word recognition processor of claim 8 wherein said parameterized representation includes a representation of the phone similarity peak location, peak height and the left and right frame locations.

Claim 12

Dependent Claim 39 words

The word recognition processor of claim 1 wherein said region count stage represents an instance of a given spoken word or phrase by the number of high similarity regions found corresponding to each of a plurality of phoneme identifiers.

Claim 13

Dependent Claim 48 words

The word recognition processor of claim 11 further comprising means for breaking said phone similarity data into a plurality of time intervals and wherein said instance of a given spoken word or phrase is represented by the number of high similarity regions in each of said time intervals.

Claim 14

Dependent Claim 27 words

The word recognition processor of claim 1 further comprising building a region count prototype corresponding to a plurality of training instances of a spoken word or phrase.

Claim 15

Dependent Claim 37 words

The word recognition processor of claim 13 wherein said region count prototype consists of statistics based on the number of high phoneme similarity regions found for each phoneme identifier in each of a plurality of time intervals.

Claim 16

Dependent Claim 38 words

The word recognition processor of claim 14 wherein said statistics comprise the mean and inverse variance of the number of said high phoneme similarity regions found for each phoneme identifier in each of said plurality of time intervals.

Claim 17

Independent Claim 135 words

A method for processing an input speech utterance for word recognition, comprising: representing the input speech utterance as a phone similarity data indicative of the correlation between the input speech utterance and predetermined phone model speech data; selecting from said phone similarity data those regions of high similarity that exceed a predetermined threshold; testing the high similarity regions against a first predetermined word prototype database using a region count procedure that selects first list of word candidates minimizing the region count distortion with respect to the input speech utterance; testing the high similarity regions of words in said first list against a second predetermined word prototype database using a target congruence procedure that selects from the first list a second list of word candidates having high similarity regions substantially congruent with the input speech utterance.

Claim 18

Independent Claim 135 words

A method for processing an input speech utterance for word recognition, comprising: representing the input speech utterance as a phone similarity data indicative of the correlation between the input speech utterance and predetermined phone model speech data; selecting from said phone similarity data those regions of high similarity that exceed a predetermined threshold; testing the high similarity regions against a first predetermined word prototype database using a region count procedure that selects first list of word candidates minimizing the region count distortion with respect to the input speech utterance; testing the high similarity regions of words in said first list against a second predetermined word prototype database using a target congruence procedure that selects from the first list a second list of word candidates having high similarity regions substantially congruent with the input speech utterance.

Claim 19

Dependent Claim 28 words

The method of claim 16 further comprising performing a fine match upon said second list of word candidates to select a single recognized word from said second list.

Claim 20

Dependent Claim 33 words

The method of claim 16 wherein said region count procedure produces a first score corresponding to the degree of fit between the input utterance and each of the first list of word candidates.

Claim 21

Dependent Claim 33 words

The method of claim 16 wherein said target congruence procedure produces a second score corresponding to the degree of fit between the input utterance and each of the second list of word candidates.

Claim 22

Dependent Claim 84 words

The method of claim 16 wherein said region count procedure produces a first score corresponding to the degree of fit between the input utterance and each of the first list of word candidates; wherein said target congruence procedure produces a second score corresponding to the degree of fit between the input utterance and each of the second list of word candidates; and further comprising combining the first and second normalized scores and selecting the word with the best score as a final word candidate.

Claim 23

Dependent Claim 19 words

The method of claim 20 wherein said combining step is performed by averaging said first and second normalized scores.

Claim 24

Dependent Claim 25 words

The method of claim 16 further comprising representing said high similarity regions as a parameter representing the high similarity regions of the phone similarity data.

Claim 25

Dependent Claim 20 words

The method of claim 22 wherein said parameters include a representation of the phone similarity peak location and peak height.

Claim 26

Dependent Claim 26 words

The method of claim 22 wherein said parameters include a representation of the phone similarity peak location, peak height and the left and right frame locations.

Claim 27

Dependent Claim 37 words

The method of claim 16 further comprising the step of: representing an instance of a given spoken word or phrase by the number of high phoneme similarity regions found for each of a plurality of phoneme identifiers.

Claim 28

Dependent Claim 42 words

The method of claim 25 further comprising breaking said phone similarity data into a plurality of time intervals and representing an instance of a given spoken word or phrase by the number of high similarity regions in each of said time intervals.

Claim 29

Dependent Claim 53 words

The method of claim 25 further comprising the step of: building a region count prototype corresponding to a spoken word or phrase that consists of statistics based on the number of high phoneme similarity regions found for each phoneme identifier in each of a plurality of time intervals in the phoneme similarity data.

Claim 30

Dependent Claim 52 words

The method of claim 27 further comprising the step of: calculating said statistics as the mean and inverse variance of the number of said high phoneme similarity regions found for each phoneme identifier, in each of a plurality of time intervals, of the training instances of the given spoken word or phrase.

Claim 31

Dependent Claim 32 words

The method of claim 27 further comprising the step of computing the recognition score for a given instance of a spoken word or phrase with respect to a given region count prototype.

Claim 32

Dependent Claim 49 words

The method of claim 29 wherein said recognition score is the Euclidean distance between: (a) the number of high phoneme similarity regions found, for each phoneme identifier, in the phoneme similarity data and (b) the mean of the number of high phoneme similarity regions found for each phoneme identifier.

Claim 33

Dependent Claim 71 words

The method of claim 29 wherein said recognition score is the Euclidean distance between: (a) the number of high phoneme similarity regions found, for each phoneme identifier, in each of a plurality of time intervals within the phoneme similarity data and (b) the mean of the number of high phoneme similarity regions, in each of said plurality of time intervals of the training instances found in the given region count prototype.

Claim 34

Dependent Claim 14 words

The method of claim 29 wherein said recognition score is a weighted Euclidean distance.

Claim 35

Dependent Claim 18 words

The method of claim 32 wherein the weight is the inverse variance from the given region count prototype.

Claim 36

Dependent Claim 20 words

The method of claim 16 further comprising: comparing the input speech utterance with each prototype to provide a recognition score.

Annuity Stage	Window Opens	Payment Deadline	Surcharge Period Until	Estimated Fee	Status
3.5 Years (Year 4)	2001-10-12	2002-04-13	2002-10-13	$2000.00	Window Closed / Lapsed
7.5 Years (Year 8)	2005-10-12	2006-04-13	2006-10-13	$3760.00	Window Closed / Lapsed
11.5 Years (Year 12)	2009-10-12	2010-04-13	2010-10-13	$7700.00	Window Closed / Lapsed

Multistage word recognizer based on reliably detected phoneme similarity regions

Abstract

Inventors

Bibliographic Details

Claims Analysis

References Cited (Backward Citations)

USPTO Annuity & Maintenance Windows

Assigned Categories & Tags

Assign Category Tag