| 摘要 |
A vector excitation coder compresses vectors by using an optimum codebook designed off line, using an initial arbitrary codebook and a set of speech training vectors exploiting codevector sparsity (i.e., by making zero all but a selected number of samples of lowest amplitude in each of N codebook vectors). A fast-search method selects a number Nc of good excitation vectors from the codebook, where Nc is much smaller than N, and uses only the Nc vectors in an exhaustive search for the best match between a perceptually weighted input vector zn, and an estimate zn derived from a codebook vector processed through long-term and short-terms filters, and a perceptual weighting filter. The zero input response of these cascaded filters is calculated and subtracted from an input speech vector sn after perceptual weighting to produce a vector rn. The codebook search operation is performed using <IMAGE> by calculating the numerator of a fast inner product and calculating the denominator by a fast inner product for each codebook vector cj, computing the right hand side of the equation once per frame, and then cross multiplying the numerators and denominators to determine if N2/D2 is less than N1/D1 by determining if N1D2>N2D1. If not N2 and D2 replace N1 and D1 in registers En and Ed.
|
