Wrapper structures of sub-vector quantization. More...
#include <subvq.h>
Public Attributes | |
arraysize_t | origsize |
int32 | n_sv |
int32 | vqsize |
int32 ** | featdim |
vector_gautbl_t * | gautbl |
int32 *** | map |
float32 * | subvec |
int32 ** | vqdist |
int32 * | gauscore |
int32 * | mgau_sl |
int32 | VQ_EVAL |
Wrapper structures of sub-vector quantization.
int32** subvq_t::featdim |
featdim[s] = Original feature dimensions in subvector s
int32* subvq_t::gauscore |
Subvq-based approx. Gaussian density scores for one mixture
Vector-quantized Gaussians table for each sub-vector
int32*** subvq_t::map |
map[i][j] = map from original codebook(i)/codeword(j) to sequence of nearest vector quantized subvector codewords; so, each map[i][j] is of length n_sv. Finally, map is LINEARIZED, so that it indexes into a 1-D array of scores rather than a 2-D array (for faster access).
int32* subvq_t::mgau_sl |
Shortlist for one mixture (based on gauscore[])
int32 subvq_t::n_sv |
#Subvectors
origsize.r = #codebooks (or states) in original model; origsize.c = max #codewords/codebook in original model.
float32* subvq_t::subvec |
Subvector extracted from feature vector
int32 subvq_t::VQ_EVAL |
int32** subvq_t::vqdist |
vqdist[i][j] = score (distance) for i-th subvector compared to j-th subvector-codeword
int32 subvq_t::vqsize |
#Codewords in each subvector quantized mean/var table