#include <stdio.h>
#include <s3types.h>
#include <logmath.h>
Go to the source code of this file.
Classes | |
| struct | gs_s |
Typedefs | |
| typedef gs_s | gs_t |
Functions | |
| int32 | gs_display (char *file, gs_t *gs) |
| gs_t * | gs_read (const char *file, logmath_t *logmath) |
| int32 | gc_compute_closest_cw (gs_t *gs, float32 *feat) |
| int32 | gs_mgau_shortlist (gs_t *gs, int32 m, int32 n, float32 *feat, int32 bst_codeid) |
Implementation of Bochierri 93's idea of Gaussian Selection. A pre-computed Gaussian selector map (not distributed in Sphinx 3 but in s3fast.tgz distributed in Arthur Chan's web page) encode the closest neighbors of a given codeword and a given GMM.
In decoding, the closest codeword of the feature is first found and the nearest neighborhood(s) for all GMMs will be retrieved from the Gaussian selector map.
The reason why this file is not opened to the public is because sub-vector quantization supersed this idea conceptually. Although our SVQ implementation do not assume multiple Gaussians to be computed given a sub-stream(sub-vector) codeword is found. It is not difficult to extend the implementation to accomplish that.
Saying this, for implementation symmetry, we should still check in the Gaussian selector implementation in this archive.
|
|
|
|
||||||||||||
|
Choose the closet Gaussian codebook to use
|
|
||||||||||||
|
display the Gaussian selector
|
|
||||||||||||||||||||||||
|
Find the short list of the Gaussian selector
|
|
||||||||||||
|
Read the Gaussian selector
|
1.3.9.1