Speech recognition device for calculating a corrected similarity partially dependent on circumstances of production of input patterns

Claim 2

A speech recognition device comprising: means for supplying an input speech signal; means for converting said input speech signal into an input pattern indicative of features of said input speech signal; a similarity measure calculating unit for calculating primary similarity measures between said input pattern and a first plurality of prepared reference patterns and for selecting a maximum value of said primary similarity measures as a provisional similarity measure; similarity measure calculating means for calculating secondary similarity measures between said input pattern and a second plurality of produced reference patterns produced in compliance with said prepared reference patterns and for selecting a maximum value of said secondary similarity measures as a reference similarity measure; a similarity measure correcting unit connected to said similarity measure calculating unit and said similarity measure calculating means for correcting said provisional similarity measure by modifying said provisional similarity measure by either a difference between said provisional similarity measure and said reference similarity measure or a ratio of said provisional similarity measure and said reference similarity measure and for outputting said corrected provisional similarity measure as a corrected similarity measure; and a normalizing unit connected to said similarity measure correcting unit for normalizing said corrected similarity measure into a normalized similarity measure by a duration of time of said input pattern.

Claim 3

A speech recognition device as claimed in claim 1, further comprising a determining unit, connected to said normalizing unit, for outputting as a recognition result when said normalized similarity measure is greater than a predetermined threshold, a signal indicative of a specific one of said prepared reference patterns that gives said provisional similarity measure.

Claim 4

A speech recognition device as claimed in claim 2, further comprising a rejection unit, connected to said normalizing unit, for outputting a reject signal indicating that said input speech signal cannot be recognized when said normalized similarity measure is not greater than said predetermined threshold value.

Claim 5

A speech recognition device comprising: means for supplying an input speech signal; means for converting said input speech signal into an input pattern indicative of features of said input speech signal; a similarity measure calculating unit for calculating primary similarity measures between said input pattern and a first plurality of prepared reference patterns and for selecting a maximum value of said primary similarity measures as a provisional similarity measure; similarity measure calculating means for calculating secondary similarity measures between said input pattern and a second plurality of produced reference patterns produced in compliance with said prepared reference patterns and for selecting a maximum value of said secondary similarity measures as a reference similarity measure; and a similarity measure correcting unit connected to said similarity measure calculating unit and said similarity measure calculating means for correcting said provisional similarity measure by modifying said provisional similarity measure by either a difference between said provisional similarity measure and said reference similarity measure or a ratio of said provisional similarity measure and said reference similarity measure and for outputting said corrected provisional similarity measure as a corrected similarity measure, said similarity measure calculating unit being a first calculating unit, said prepared reference patterns being stored reference patterns stored in said first calculating unit, wherein said similarity measure calculating means comprises: a recognition unit memory for storing recognition units of said stored reference patterns; and a second calculating unit connected to said recognition unit memory and supplied with said input pattern for calculating said secondary similarity measures between said input pattern and concatenated reference patterns of a second plurality in number and for selecting a maximum value of said secondary similarity measures as said reference similarity measure, said concatenated reference patterns serving as said produced reference patterns and being concatenations of selected units which are selected from said recognition units in compliance with said stored reference patterns.

Claim 6

A speech recognition device comprising: means for supplying an input speech signal; means for converting said input speech signal into an input pattern indicative of features of said input speech signal; a similarity measure calculating unit for calculating primary similarity measures between said input pattern and a first plurality of prepared reference patterns and for selecting a maximum value of said primary similarity measures as a provisional similarity measure; similarity measure calculating means for calculating secondary similarity measures between said input pattern and a second plurality of produced reference patterns produced in compliance with said prepared reference patterns and for selecting a maximum value of said secondary similarity measures as a reference similarity measure; and a similarity measure correcting unit connected to said similarity measure calculating unit and said similarity measure calculating means for correcting said provisional similarity measure by modifying said provisional similarity measure by either a difference between said provisional similarity measure and said reference similarity measure or a ratio of said provisional similarity measure and said reference similarity measure and for outputting said corrected provisional similarity measure as a corrected similarity measure, said similarity measure calculating unit being a first calculating unit, said prepared reference patterns being stored reference patterns stored in said first calculating unit, wherein said similarity measure calculating means comprises: a recognition unit memory for storing recognition units of said stored reference patterns; and a second calculating unit connected to said recognition unit memory and supplied with said input pattern for calculating said secondary similarity measures between said input pattern and concatenated reference patterns of a second plurality in number and for selecting a maximum value of said secondary similarity measures as said reference similarity measure, said concatenated reference patterns serving as said produced reference patterns and being concatenations of selected units which are selected from said recognition units in compliance with said stored reference patterns.

Claim 7

A speech recognition device as claimed in claim 4, further comprising a normalizing unit, connected to said similarity correcting unit, for normalizing said corrected similarity measure into a normalized similarity measure by a duration of time of said input pattern.

Claim 8

A speech recognition device as claimed in claim 5, further comprising a determining unit, connected to said normalizing unit for outputting, as a recognition result of said input speech signal when said normalized similarity measure is greater than said predetermined threshold value, a signal indicative of a specific one of said prepared reference patterns that gives said provisional similarity measure.

Claim 9

A speech recognition device as claimed in claim 6, further comprising a rejection unit, connected to said normalizing unit, for outputting a reject signal indicating that said input speech signal cannot be recognized when said normalized similarity measure is not greater than said predetermined threshold value.

Claim 10

A speech recognition device comprising: means for supplying an input speech signal; means for converting said input speech signal into an input pattern indicative of features of said input speech signal; a similarity measure calculating unit for calculating primary similarity measures between said input pattern and a first plurality of prepared reference patterns and for selecting a maximum value of said primary similarity measures as a provisional similarity measure; similarity measure calculating means for calculating secondary similarity measures between said input pattern and a second plurality of produced reference patterns produced in compliance with said prepared reference patterns and for selecting a maximum value of said secondary similarity measures as a reference similarity measure; and a similarity measure correcting unit connected to said similarity measure calculating unit and said similarity measure calculating means for correcting said provisional similarity measure by modifying said provisional similarity measure by either a difference between said provisional similarity measure and said reference similarity measure or a ratio of said provisional similarity measure and said reference similarity measure and for outputting said corrected provisional similarity measure as a corrected similarity measure, said input pattern being represented by a time sequence which is divisible into time sequence frames and comprises input feature vectors, said prepared reference patterns being stored reference patterns stored in said similarity measure calculating unit, each of said stored reference patterns being represented by a time sequence of stored feature vectors, wherein said similarity measure calculating means comprises: an inter-vector similarity calculating unit connected to said similarity measure calculating unit and supplied with said input pattern for calculating inter-vector similarity measures between the input feature vectors of said time sequence frames and the stored feature vectors of said stored reference patterns with the inter-vector similarity measures of a first plurality in number calculated relative to each of said time sequence frames and for selecting maximum values of the inter-vector similarity measures calculated relative to said time sequence frames, respectively; and an accumulating unit connected to said inter-vector similarity calculating unit for accumulating said maximum values into an accumulation for use as said reference similarity measure.

Claim 11

A speech recognition device comprising: means for supplying an input speech signal; means for converting said input speech signal into an input pattern indicative of features of said input speech signal; a similarity measure calculating unit for calculating primary similarity measures between said input pattern and a first plurality of prepared reference patterns and for selecting a maximum value of said primary similarity measures as a provisional similarity measure; similarity measure calculating means for calculating secondary similarity measures between said input pattern and a second plurality of produced reference patterns produced in compliance with said prepared reference patterns and for selecting a maximum value of said secondary similarity measures as a reference similarity measure; and a similarity measure correcting unit connected to said similarity measure calculating unit and said similarity measure calculating means for correcting said provisional similarity measure by modifying said provisional similarity measure by either a difference between said provisional similarity measure and said reference similarity measure or a ratio of said provisional similarity measure and said reference similarity measure and for outputting said corrected provisional similarity measure as a corrected similarity measure, said input pattern being represented by a time sequence which is divisible into time sequence frames and comprises input feature vectors, said prepared reference patterns being stored reference patterns stored in said similarity measure calculating unit, each of said stored reference patterns being represented by a time sequence of stored feature vectors, wherein said similarity measure calculating means comprises: an inter-vector similarity calculating unit connected to said similarity measure calculating unit and supplied with said input pattern for calculating inter-vector similarity measures between the input feature vectors of said time sequence frames and the stored feature vectors of said stored reference patterns with the inter-vector similarity measures of a first plurality in number calculated relative to each of said time sequence frames and for selecting maximum values of the inter-vector similarity measures calculated relative to said time sequence frames, respectively; and an accumulating unit connected to said inter-vector similarity calculating unit for accumulating said maximum values into an accumulation for use as said reference similarity measure.

Claim 12

A speech recognition device as claimed in claim 8, further comprising a normalizing unit, connected to said similarity correcting unit, for normalizing said corrected similarity measure into a normalized similarity measure by a duration of time of said input pattern.

Claim 13

A speech recognition device as claimed in claim 9, further comprising a determining unit, connected to said normalizing unit for outputting, as a recognition result of said input speech signal when said normalized similarity measure is greater than said predetermined threshold value, a signal indicative of a specific one of said prepared reference patterns that gives said provisional similarity measure.

Claim 14

A speech recognition device as claimed in claim 10, further comprising a rejection unit, connected to said normalizing unit, for outputting a reject signal indicating that said input speech signal cannot be recognized when said normalized similarity measure is not greater than said predetermined threshold value.

Claim 15

A speech recognition device comprising: means for supplying an input speech signal; means for converting said input speech signal into an input pattern indicative of features of said input speech signal; a similarity measure calculating unit for calculating primary similarity measures between said input pattern and a first plurality of prepared reference patterns and for selecting a maximum value of said primary similarity measures as a provisional similarity measure; similarity measure calculating means for calculating secondary similarity measures between said input pattern and a second plurality of produced reference patterns produced in compliance with said prepared reference patterns and for selecting a maximum value of said secondary similarity measures as a reference similarity measure; and a similarity measure correcting unit connected to said similarity measure calculating unit and said similarity measure calculating means for correcting said provisional similarity measure by modifying said provisional similarity measure by either a difference between said provisional similarity measure and said reference similarity measure or a ratio of said provisional similarity measure and said reference similarity measure and for outputting said corrected provisional similarity measure as a corrected similarity measure, said similarity measure calculating unit comprising a first recognition unit memory for storing primary recognition units of said prepared reference patterns and a first calculating unit connected to said first recognition unit memory and supplied with said input pattern for calculating said primary similarity measures between said input pattern and primary concatenated reference patterns of a first plurality in number and for selecting a maximum value of said primary similarity measures as said provisional similarity measure, said primary concatenated reference patterns serving as said prepared reference patterns and being concatenations of primary selected units which are selected from said primary recognition units in compliance with said prepared reference patterns, wherein: said similarity measure calculating means comprises: a second recognition unit memory for storing secondary recognition units selected in compliance with said prepared reference patterns; and a second calculating unit connected to said second recognition unit memory and supplied with said input pattern for calculating said secondary similarity measures between said input pattern and secondary concatenated reference patterns of said second plurality in number and for selecting a maximum value of said secondary similarity measures as said reference similarity measure, said secondary concatenated reference patterns serving as said produced reference patterns and being concatenations of secondary selected units which are selected from said secondary recognition units in compliance with said prepared reference patterns; said input pattern becoming a succession of segment patterns corresponding to particular units concatenated among said primary selected units into one of said primary concatenated reference patterns that gives said provisional similarity measure, said similarity measure correcting unit comprising: a unit similarity calculating unit connected to said first calculating unit for calculating unit similarity measures between said segment patterns and said particular units; and a unit similarity correcting unit connected to said unit similarity calculating unit and said second calculating unit for dividing said reference similarity measure into interval similarity measures in correspondence to said particular units and for correcting each of said unit similarity measures into said corrected similarity measure by one of said interval similarity measures that corresponds to one of said particular units, said one of the particular units being used in calculating said each of the unit similarity measures.

Claim 16

A speech recognition device comprising: means for supplying an input speech signal; means for converting said input speech signal into an input pattern indicative of features of said input speech signal; a similarity measure calculating unit for calculating primary similarity measures between said input pattern and a first plurality of prepared reference patterns and for selecting a maximum value of said primary similarity measures as a provisional similarity measure; similarity measure calculating means for calculating secondary similarity measures between said input pattern and a second plurality of produced reference patterns produced in compliance with said prepared reference patterns and for selecting a maximum value of said secondary similarity measures as a reference similarity measure; and a similarity measure correcting unit connected to said similarity measure calculating unit and said similarity measure calculating means for correcting said provisional similarity measure by modifying said provisional similarity measure by either a difference between said provisional similarity measure and said reference similarity measure or a ratio of said provisional similarity measure and said reference similarity measure and for outputting said corrected provisional similarity measure as a corrected similarity measure, said similarity measure calculating unit comprising a first recognition unit memory for storing primary recognition units of said prepared reference patterns and a first calculating unit connected to said first recognition unit memory and supplied with said input pattern for calculating said primary similarity measures between said input pattern and primary concatenated reference patterns of a first plurality in number and for selecting a maximum value of said primary similarity measures as said provisional similarity measure, said primary concatenated reference patterns serving as said prepared reference patterns and being concatenations of primary selected units which are selected from said primary recognition units in compliance with said prepared reference patterns, wherein: said similarity measure calculating means comprises: a second recognition unit memory for storing secondary recognition units selected in compliance with said prepared reference patterns; and a second calculating unit connected to said second recognition unit memory and supplied with said input pattern for calculating said secondary similarity measures between said input pattern and secondary concatenated reference patterns of said second plurality in number and for selecting a maximum value of said secondary similarity measures as said reference similarity measure, said secondary concatenated reference patterns serving as said produced reference patterns and being concatenations of secondary selected units which are selected from said secondary recognition units in compliance with said prepared reference patterns; said input pattern becoming a succession of segment patterns corresponding to particular units concatenated among said primary selected units into one of said primary concatenated reference patterns that gives said provisional similarity measure, said similarity measure correcting unit comprising: a unit similarity calculating unit connected to said first calculating unit for calculating unit similarity measures between said segment patterns and said particular units; and a unit similarity correcting unit connected to said unit similarity calculating unit and said second calculating unit for dividing said reference similarity measure into interval similarity measures in correspondence to said particular units and for correcting each of said unit similarity measures into said corrected similarity measure by one of said interval similarity measures that corresponds to one of said particular units, said one of the particular units being used in calculating said each of the unit similarity measures.

Claim 17

A speech recognition device as claimed in claim 12, further comprising a normalizing unit, connected to said similarity correcting unit, for normalizing said corrected similarity measure into a normalized similarity measure by a duration of time of one of said segment patterns that is used in calculating said each of the unit similarity measures.

Claim 18

A speech recognition device as claimed in claim 13, further comprising a determining unit, connected to said normalizing unit for, when said normalized similarity measure is not greater than said predetermined threshold value, outputting as a recognition result of one of said segment patterns that gives said each of the unit similarity measures, a signal indicative of one of secondary selected units that is used in said secondary concatenated reference patterns and is related to said one of the interval similarity measures.

Claim 19

A speech recognition device as claimed in claim 14, further comprising a rejection unit connected to said normalizing unit for outputting a reject signal indicating that said input speech signal cannot be recognized when said normalized similarity measure is not greater than said predetermined threshold value.

Claim 20

A speech recognition device comprising: means for supplying an input speech signal; means for converting said input speech signal into an input pattern indicative of features of said input speech signal; a similarity measure calculating unit for calculating primary similarity measures between said input pattern and a first plurality of prepared reference patterns and for selecting a maximum value of said primary similarity measures as a provisional similarity measure; similarity measure calculating means for calculating secondary similarity measures between said input pattern and a second plurality of produced reference patterns produced in compliance with said prepared reference patterns and for selecting a maximum value of said secondary similarity measures as a reference similarity measure; and a similarity measure correcting unit connected to said similarity measure calculating unit and said similarity measure calculating means for correcting said provisional similarity measure by modifying said provisional similarity measure by either a difference between said provisional similarity measure and said reference similarity measure or a ratio of said provisional similarity measure and said reference similarity measure and for outputting said corrected provisional similarity measure as a corrected similarity measure, said input pattern being represented by a time sequence of input feature vectors, said similarity measure calculating unit comprising a recognition unit memory for storing recognition units of said prepared reference patterns with each of said recognition units represented by a time sequence of stored feature vectors and an elementary calculating unit connected to said recognition unit memory and supplied with said input pattern for calculating, in connection with said input feature vectors and the stored feature vectors of said recognition units, said primary similarity measures between said input pattern and concatenated reference patterns of a first plurality in number and for selecting a maximum value of said primary similarity measures as said provisional similarity measure, said concatenated reference patterns serving as said prepared reference patterns and being concatenations of selected units which are selected from said recognition units in compliance with said prepared reference patterns, the time sequence of said input feature vectors being divisible into time sequence frames; said similarity measure calculating means comprising: an inter-vector similarity calculating unit connected to said recognition unit memory and supplied with said input pattern for calculating inter-vector similarity measures between the input feature vectors of said time sequence frames and the stored feature vectors of said recognition units with the inter-vector similarity measures of a second plurality in number calculated relative to each of said time sequence frames and for selecting maximum values of the inter-vector similarity measures which are calculated relative to said time sequence frames, respectively; and an accumulating unit connected to said inter-vector similarity calculating unit for accumulating said maximum values into an accumulation for use as said reference similarity measure; said input pattern becoming a succession of segment patterns corresponding to particular units concatenated among selected units into one of concatenated reference patterns that gives said provisional similarity measure, said similarity measure correcting unit comprising: a unit similarity calculating unit connected to said elementary calculating unit for calculating unit similarity measures between said segment patterns and said particular units; and a unit similarity correcting unit connected to said unit similarity calculating unit and said accumulating unit for dividing said reference similarity measure into interval similarity measures in correspondence to said particular units and for correcting each of said unit similarity measures into said corrected similarity measure by one of said interval similarity measures that corresponds to one of said particular units, said one of the particular units being used in calculating said each of the unit similarity measures.

Claim 21

A speech recognition device comprising: means for supplying an input speech signal; means for converting said input speech signal into an input pattern indicative of features of said input speech signal; a similarity measure calculating unit for calculating primary similarity measures between said input pattern and a first plurality of prepared reference patterns and for selecting a maximum value of said primary similarity measures as a provisional similarity measure; similarity measure calculating means for calculating secondary similarity measures between said input pattern and a second plurality of produced reference patterns produced in compliance with said prepared reference patterns and for selecting a maximum value of said secondary similarity measures as a reference similarity measure; and a similarity measure correcting unit connected to said similarity measure calculating unit and said similarity measure calculating means for correcting said provisional similarity measure by modifying said provisional similarity measure by either a difference between said provisional similarity measure and said reference similarity measure or a ratio of said provisional similarity measure and said reference similarity measure and for outputting said corrected provisional similarity measure as a corrected similarity measure, said input pattern being represented by a time sequence of input feature vectors, said similarity measure calculating unit comprising a recognition unit memory for storing recognition units of said prepared reference patterns with each of said recognition units represented by a time sequence of stored feature vectors and an elementary calculating unit connected to said recognition unit memory and supplied with said input pattern for calculating, in connection with said input feature vectors and the stored feature vectors of said recognition units, said primary similarity measures between said input pattern and concatenated reference patterns of a first plurality in number and for selecting a maximum value of said primary similarity measures as said provisional similarity measure, said concatenated reference patterns serving as said prepared reference patterns and being concatenations of selected units which are selected from said recognition units in compliance with said prepared reference patterns, the time sequence of said input feature vectors being divisible into time sequence frames; said similarity measure calculating means comprising: an inter-vector similarity calculating unit connected to said recognition unit memory and supplied with said input pattern for calculating inter-vector similarity measures between the input feature vectors of said time sequence frames and the stored feature vectors of said recognition units with the inter-vector similarity measures of a second plurality in number calculated relative to each of said time sequence frames and for selecting maximum values of the inter-vector similarity measures which are calculated relative to said time sequence frames, respectively; and an accumulating unit connected to said inter-vector similarity calculating unit for accumulating said maximum values into an accumulation for use as said reference similarity measure; said input pattern becoming a succession of segment patterns corresponding to particular units concatenated among selected units into one of concatenated reference patterns that gives said provisional similarity measure, said similarity measure correcting unit comprising: a unit similarity calculating unit connected to said elementary calculating unit for calculating unit similarity measures between said segment patterns and said particular units; and a unit similarity correcting unit connected to said unit similarity calculating unit and said accumulating unit for dividing said reference similarity measure into interval similarity measures in correspondence to said particular units and for correcting each of said unit similarity measures into said corrected similarity measure by one of said interval similarity measures that corresponds to one of said particular units, said one of the particular units being used in calculating said each of the unit similarity measures.

Claim 22

A speech recognition device as claimed in claim 16, further comprising a normalizing unit, connected to said similarity correcting unit, for normalizing said corrected similarity measure into a normalized similarity measure by a duration of time of one of said segment patterns that is used in calculating said each of the unit similarity measures.

Claim 23

A speech recognition device as claimed in claim 17, further comprising a determining unit, connected to said normalizing unit, for outputting as a recognition result when said normalized similarity measure is greater than a predetermined threshold, a signal indicative of a specific one of said prepared reference patterns that gives said provisional similarity measure.

Claim 24

A speech recognition device as claimed in claim 18, further comprising a rejection unit connected to said normalizing unit for outputting a reject signal indicating that said input speech signal cannot be recognized when said normalized similarity measure is not greater than said predetermined threshold value.