초록 ▼

A method, apparatus, and computer software are provided whereby to associate a readily user-interpretable reason with an output of a supervised training data classifier. Reasons are associated with one or more members of a sequence of training vectors and subsequently associated, in operation, with

A method, apparatus, and computer software are provided whereby to associate a readily user-interpretable reason with an output of a supervised training data classifier. Reasons are associated with one or more members of a sequence of training vectors and subsequently associated, in operation, with a given classifier input by comparing the classifier input vector with training sequence vectors. A measure of confidence in the selected reasons is derived by comparing the classifier input vector with the corresponding training inputs with which the selected reasons were associated and calculating a measure of their closeness.

대표청구항 ▼

A method, apparatus, and computer software are provided whereby to associate a readily user-interpretable reason with an output of a supervised training data classifier. Reasons are associated with one or more members of a sequence of training vectors and subsequently associated, in operation, with

A method, apparatus, and computer software are provided whereby to associate a readily user-interpretable reason with an output of a supervised training data classifier. Reasons are associated with one or more members of a sequence of training vectors and subsequently associated, in operation, with a given classifier input by comparing the classifier input vector with training sequence vectors. A measure of confidence in the selected reasons is derived by comparing the classifier input vector with the corresponding training inputs with which the selected reasons were associated and calculating a measure of their closeness. forming a piece-wise constant speech power spectral density estimate by averaging power densities within frequency bin segments of said speech high frequency-resolution power spectral density estimate; determining a high frequency-resolution residual echo power spectral density estimate from an echo signal block; forming a piece-wise constant residual echo power spectral density estimate by averaging power densities within frequency bin segments of said high frequency-resolution residual echo power spectral density estimate; determining a piece-wise constant digital filter transfer function using said piece-wise constant speech power spectral density estimate and said piece-wise constant residual echo power spectral density estimate; wherein said step of forming a piece-wise constant speech power spectral density estimate includes adapting the length of individual segments to the shape of said high frequency-resolution speech power spectral density estimate; and wherein said step of forming a piece-wise constant residual echo power spectral density estimate includes using the same segments for both said high frequency-resolution speech power spectral density estimate and said high frequency-resolution residual echo power spectral density estimate. 6. The method of claim 5, including the step of centering segments on local maxima of said high frequency-resolution speech power spectral density estimate. 7. The method of claim 6, including the step of increasing said segment length for high frequencies in accordance with the human auditory system. 8. The method of claim 5, including the step of increasing said segment length for high frequencies in accordance with the human auditory system. 9. A digital filter design method, including the steps of: determining a high frequency-resolution power spectral density estimate from an input signal block; forming a piece-wise constant power spectral density estimate by averaging power densities within frequency bin segments of said high frequency resolution power spectral density estimate; determining a piece-wise constant digital filter transfer function using said piece-wise constant power spectral density estimate; wherein said step of forming a piece-wise constant power spectral density estimate includes adapting the length of individual segments to the shape of said high frequency-resolution power spectral density estimate. 10. The method of claim 9, including the step of centering segments on local maxima of said high frequency-resolution power spectral density estimates. 11. A digital filter design apparatus for noise suppression by spectral subtraction, including: means for determining a high frequency-resolution noisy speech power spectral density estimate from a noisy speech signal block; means for forming a piece-wise constant noisy speech power spectral density estimate by averaging power densities within frequency bin segments of said noisy speech high frequency-resolution power spectral density estimate; means for determining a high frequency-resolution background noise power spectral density estimate from a background noise signal block; means for forming a piece-wise constant background noise power spectral density estimate by averaging power densities within frequency bin segments of said high frequency-resolution background noise power spectral density estimate; means for determining a piece-wise constant digital filter transfer function using spectral subtraction based on said piece-wise constant noisy speech power spectral density estimate and said piece-wise constant back ground noise power spectral density estimate; means for adapting the length of individual segments to the shape of said high frequency-resolution noisy speech power spectral density estimate; and means for using the same segments for both said high frequency resolution noisy speech power spectral density estimate and said high frequency-resolution background noise power spect