Method and recognizer for recognizing a sampled sound signal in noise

Claim 3

A method according to claim 1 wherein each of the normalized feature vectors is found according to an equation: v˜.sub.ij =(v.sub.ij -min.sub.j)/(max.sub.j -min.sub.j) where v˜ ij is one of the normalized feature values of an analysis frame "i" of the number of analysis frames and a frequency band "j" of the plurality of frequency bands; v ij is one of the feature values at the analysis frame "i" and the frequency band "j"; min j is the minimum feature value for the "j" frequency band; and max j is the maximum feature value for the "j" frequency band.

Claim 4

A method according to claim 1 wherein the step of computing feature values comprises: dividing a power spectrum of the acoustic sound signal into the plurality of frequency bands; weighting the power spectrum within each of the plurality of frequency bands according to a weighting function to obtain filter outputs; and calculating the feature values from the filter outputs.

Claim 5

A method according to claim 3 wherein the power spectrum is weighted in a frequency domain according to a raised cosine profile.

Claim 6

A method according to claim 4 wherein the raised cosine profile calculates the filter outputs according to an equation: f.sub.ij =ΣP.sub.i (ω)B.sub.j (ω) where f ij is one of the filter outputs at an analysis frame "i" of the number of analysis frames and a frequency band "j" of the plurality of frequency bands, P i (ω) is a power spectrum value for the analysis frame "i" at frequency ω, B j (ω) is a band-pass filter weighting function, or frequency response, for the frequency band "j" at frequency ω, and Σ represents a summation operation over all frequencies ω.

Claim 7

A method according to claim 1 wherein the step of comparing the normalized feature vectors with sets of template feature vectors to determine the template feature vectors that most resemble the normalized feature vectors includes comparing the normalized feature vectors with the sets of template feature vectors in a dynamic time warping process.

Claim 8

A method according to claim 3 wherein the step of calculating the feature values from the filter outputs comprises taking a log of each of the filter outputs.

Claim 9

A method according to claim 3 wherein the power spectrum is divided into the plurality of frequency bands according to a Mel or Bark scale to approximate a frequency response of a human ear.

Claim 10

A sound recognizer comprising: a feature vector device which computes feature values for a number of analysis frames of an acoustic signal input to the sound recognizer, the feature values computed for a plurality of frequency bands; a min/max device coupled to the feature vector device to determine in a logarithmic domain which of the feature values within a respective one of the plurality of frequency bands is a minimum feature value over time and which of the feature values within the respective one of the plurality of frequency bands is a maximum feature value over time for each of the plurality of frequency bands across all the number of analysis frames; a normalizer coupled to the min/max device which compares, within each of the plurality of frequency bands, each of the feature values with the minimum feature value and the maximum feature value of the respective one of the plurality of frequency bands to obtain normalized feature values, normalized over time within each of the frequency bands, wherein all of the normalized feature values for a given one of the analysis frames define one of a plurality of normalized feature vectors; and a comparator coupled to the normalizer which compares the plurality of normalized feature vectors with sets of template feature vectors to determine template feature vectors that most resemble the plurality of normalized feature vectors, the comparator outputting the template feature vectors that most resemble the plurality of normalized feature vectors.

Claim 11

A sound recognizer comprising: a feature vector device which computes feature values for a number of analysis frames of an acoustic signal input to the sound recognizer, the feature values computed for a plurality of frequency bands; a min/max device coupled to the feature vector device to determine in a logarithmic domain which of the feature values within a respective one of the plurality of frequency bands is a minimum feature value over time and which of the feature values within the respective one of the plurality of frequency bands is a maximum feature value over time for each of the plurality of frequency bands across all the number of analysis frames; a normalizer coupled to the min/max device which compares, within each of the plurality of frequency bands, each of the feature values with the minimum feature value and the maximum feature value of the respective one of the plurality of frequency bands to obtain normalized feature values, normalized over time within each of the frequency bands, wherein all of the normalized feature values for a given one of the analysis frames define one of a plurality of normalized feature vectors; and a comparator coupled to the normalizer which compares the plurality of normalized feature vectors with sets of template feature vectors to determine template feature vectors that most resemble the plurality of normalized feature vectors, the comparator outputting the template feature vectors that most resemble the plurality of normalized feature vectors.

Claim 12

A sound recognizer according to claim 9 wherein the normalizer computes each of the plurality of normalized feature vectors according to the equation: v˜.sub.ij =(v.sub.ij -min.sub.j)/(max.sub.j -min.sub.j) where v˜ ij is one of the normalized feature values of an analysis frame "i" of the number of analysis frames and a frequency bands at "j" of the plurality of frequency bands; v ij is one of the feature values at the analysis frame "i" and the frequency band "j"; min j is the minimum feature value for the "j" frequency band; and max j is the maximum feature value for the "j" frequency band.

Claim 13

A sound recognizer according to claim 9 wherein the comparator is coupled to a template feature vector library, the template feature vector library containing the sets of template feature vectors.

Claim 16

A method according to claim 12 wherein the step of computing a plurality of feature values comprises: dividing a power spectrum of the acoustic sound signal into the plurality of frequency bands; weighting the power spectrum within each of the plurality of frequency bands according to a weighting function to obtain filter outputs; and calculating the plurality of feature values from the filter outputs.

Claim 17

A method according to claim 13 wherein the power spectrum is weighted in a frequency domain according to a raised cosine profile.

Claim 18

A method according to claim 14 wherein the raised cosine profile calculates the filter outputs according to an equation: f.sub.ij =ΣP.sub.i (ω)B.sub.j (ω) where f ij is one of the filter outputs at an analysis frame "i" of the number of analysis frames and a frequency band "j" of the plurality of frequency bands, P i (ω) is a power spectrum value for the analysis frame "i" at frequency ω, B j (ω) is a band-pass filter weighting function, or frequency response, for the frequency band "j" at frequency ω, and Σ represents a summation operation over all frequencies ω.

Claim 19

A method according to claim 12 wherein the step of comparing the normalized feature vectors with sets of template feature vectors to determine template feature vectors that most resemble the normalized feature vectors includes comparing the normalized feature vectors with the sets of template feature vectors in a dynamic time warping process.

Claim 20

A method according to claim 13 wherein the step of calculating the plurality of feature values from the filter outputs comprises taking a log of each of the filter outputs.

Claim 21

A method according to claim 13 wherein the power spectrum is divided into the plurality of frequency bands according to a Mel or Bark scale to approximate a frequency response of a human ear.

Claim 22

A sound recognizer comprising: a feature vector device which computes feature values for a number of analysis frames of an acoustic signal input to the sound recognizer, the feature values computed for a plurality of frequency bands; a min/max device coupled to the feature vector device to determine in a logarithmic domain which of the feature values within a respective one of the plurality of frequency bands is a minimum feature value over time and which of the feature values within the respective one of the plurality of frequency bands is a maximum feature value over time for each of the plurality of frequency bands across all the number of analysis frames; a normalizer coupled to the min/max device which compares each of the feature values within each of the plurality of frequency bands with the minimum feature value and the maximum feature value of the respective one of the plurality of frequency bands to obtain normalized feature values, wherein all of the normalized feature values for a given one of the analysis frames define one of a plurality of normalized feature vectors, each of the normalized feature values found according to an equation: v˜.sub.ij =(v.sub.ij -min.sub.j)/(max.sub.j -min.sub.j) where v˜ ij is one of the normalized feature values of an analysis frame "i" of the number of analysis frames and a frequency bands at "j" of the plurality of frequency bands, v ij is one of the feature values at the analysis frame "i" and the frequency band "j", min j is the minimum feature value for the "j" frequency band, and max j is the maximum feature value for the "j" frequency band; and a comparator coupled to the normalizer which compares the plurality of normalized feature vectors with sets of template feature vectors to determine template feature vectors that most resemble the plurality of normalized feature vectors, the comparator outputting the template feature vectors that most resemble the plurality of normalized feature vectors.

Claim 23

A sound recognizer comprising: a feature vector device which computes feature values for a number of analysis frames of an acoustic signal input to the sound recognizer, the feature values computed for a plurality of frequency bands; a min/max device coupled to the feature vector device to determine in a logarithmic domain which of the feature values within a respective one of the plurality of frequency bands is a minimum feature value over time and which of the feature values within the respective one of the plurality of frequency bands is a maximum feature value over time for each of the plurality of frequency bands across all the number of analysis frames; a normalizer coupled to the min/max device which compares each of the feature values within each of the plurality of frequency bands with the minimum feature value and the maximum feature value of the respective one of the plurality of frequency bands to obtain normalized feature values, wherein all of the normalized feature values for a given one of the analysis frames define one of a plurality of normalized feature vectors, each of the normalized feature values found according to an equation: v˜.sub.ij =(v.sub.ij -min.sub.j)/(max.sub.j -min.sub.j) where v˜ ij is one of the normalized feature values of an analysis frame "i" of the number of analysis frames and a frequency bands at "j" of the plurality of frequency bands, v ij is one of the feature values at the analysis frame "i" and the frequency band "j", min j is the minimum feature value for the "j" frequency band, and max j is the maximum feature value for the "j" frequency band; and a comparator coupled to the normalizer which compares the plurality of normalized feature vectors with sets of template feature vectors to determine template feature vectors that most resemble the plurality of normalized feature vectors, the comparator outputting the template feature vectors that most resemble the plurality of normalized feature vectors.

Claim 24

A voice recognizer according to claim 19 wherein the comparator is coupled to a template feature vector library, the template feature vector library containing the template feature vectors.

Claim 25

A sound recognizer comprising: a feature vector device which computes feature values for a number of analysis frames of an acoustic signal input to the sound recognizer, the feature values computed for a plurality of frequency bands; a min/max device coupled to the feature vector device to determine in a logarithmic domain which of the feature values within a respective one of the plurality of frequency bands is a minimum feature value and which of the feature values within the respective one of the plurality of frequency bands is a maximum feature value for each of the plurality of frequency bands across all the number of analysis frames; a normalizer coupled to the min/max device which compares each of the feature values within each of the plurality of frequency bands with the minimum feature value and the maximum feature value of the respective one of the plurality of frequency bands to obtain normalized feature values, wherein all of the normalized feature values for a given one of the analysis frames defines one of a plurality of normalized feature vectors; and a comparator coupled to the normalizer which compares the plurality of normalized feature vectors with sets of template feature vectors to determine template feature vectors that most resemble the plurality of normalized feature vectors, the comparator outputting the template feature vectors that most resemble the plurality of normalized feature vectors, wherein the min/max device has a weighted minimum feature value and a weighted maximum feature value for frequency band "j" for all analysis frames, 0≦i≦n, where the minimum and maximum feature values are computed as follows: ##EQU3## where weight k! is a weighting function having a central point having a value greater than or equal to all other weight values.

Claim 26

A sound recognizer comprising: a feature vector device which computes feature values for a number of analysis frames of an acoustic signal input to the sound recognizer, the feature values computed for a plurality of frequency bands; a min/max device coupled to the feature vector device to determine in a logarithmic domain which of the feature values within a respective one of the plurality of frequency bands is a minimum feature value and which of the feature values within the respective one of the plurality of frequency bands is a maximum feature value for each of the plurality of frequency bands across all the number of analysis frames; a normalizer coupled to the min/max device which compares each of the feature values within each of the plurality of frequency bands with the minimum feature value and the maximum feature value of the respective one of the plurality of frequency bands to obtain normalized feature values, wherein all of the normalized feature values for a given one of the analysis frames defines one of a plurality of normalized feature vectors; and a comparator coupled to the normalizer which compares the plurality of normalized feature vectors with sets of template feature vectors to determine template feature vectors that most resemble the plurality of normalized feature vectors, the comparator outputting the template feature vectors that most resemble the plurality of normalized feature vectors, wherein the min/max device has a weighted minimum feature value and a weighted maximum feature value for frequency band "j" for all analysis frames, 0≦i≦n, where the minimum and maximum feature values are computed as follows: ##EQU3## where weight k! is a weighting function having a central point having a value greater than or equal to all other weight values.

Claim 27

A sound recognizer according to claim 21 where ##EQU4##