Tuesday, December 14, 2010

Honda's 7,853,539 : Discriminating speech and non-speech

The first claim of Honda's '539 patent:

A computer implemented methodology for sound discrimination, comprising: receiving a training set of sound waveforms; generating a cortical representation of the training set; computing a matrix decomposition using the cortical representation of the training set including computing a singular value decomposition (SVD) using the cortical representation of the training set, wherein the SVD is the matrix decomposition; receiving a plurality of regularization parameters; computing coefficients for each regularization parameter using the matrix decomposition; computing a leave-one-out (LOO) error for each of the regularization parameters; and selecting the regularization parameter with the lowest LOO error.

0 Comments:

Post a Comment

<< Home