SPRAAK
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Groups Pages
Functions
sspmod_mvg.c File Reference

Evaluate a set of GMMs. More...

Functions

void spr_mvg_eval_free (SprSspInfo *Info)
 
int spr_mvg_eval_setup (SprSspInfo *Info, const char **descript, void *aux_info)
 
int spr_mvg_eval_process (SprSspInfo *Info, const void *frame_in, void *frame_out)
 

Detailed Description

Evaluate a set of GMMs.

Evaluate a set of Gaussians of a set of Gaussian Mixure Models (GMMs) with shared Gaussians and return the (normalized) Gaussian likelihoods or the (normalized) GMM likelihoods (log10 or linear).


[mvg_eval]
%
Evaluate Gaussians or GMMs and return their (log10) likelihoods.
load <acmod.mvg> [acmod.hmm] [acmod.sel]
Load a set of Gaussians (only <acmod.mvg> specified) or a set of Gaussian mixtures with shared Gaussians (<acmod.hmm> and <acmod.sel> also specified).
options <options>(LIN;NORM;NOFLOOR;LIKELIHOOD)
options (options are seperated by a ';' character):
log=<int:val>(0)
indicate whether the input/output is / must be stored as linear values of log values
LIN
indicate that the input/output is / must be stored as linear values
LOG
indicate that the input/output is / must be stored as log values
LOG10(NULL)
indicate that the input/output is / must be stored as log10 values
scale=<F32:val>(-1)
the input/output values are / must be multiplied with this factor (log values) or raisen to that power (linear)
scale_inv=<F32:val>(-1)
1.0/<scale>, i.e. multiply with this factor (log values) or raise to that power (linear) to obtain natural log / linear values
norm=<int:val>(1)
indicate whether the input/output is/must be normalized (1) or not (0)
NORM
indicate that the input/output is / must be normalized
NONORM
indicate that the input/output is not / must not be normalized
floor=<int:val>(0)
indicate whether the input/output is / must be floored (1) or not (0)
FLOOR
indicate that the input/output is / must be floored
NOFLOOR
indicate that the input/output is not / must not be floored
postprob=<int:val>(0)
indicate whether the inputs/outputs are (scaled) posteriors (1) or (scaled) likelihoods (0)
POSTPROB
indicate that the inputs/outputs are (scaled) posteriors
LIKELIHOOD
indicate that the inputs/outputs are (scaled) likelihoods
tot_prob=<int:val>(0)
indicate whether the total (log) likelihood is / must be calculated or not
TOT_PROB(NULL)
indicate whether the total (log) likelihood is / must be calculated
entropy=<int:val>(0)
indicate whether the entropy is / must be calculated or not
ENTROPY(NULL)
the entropy is / must be calculated
Various options to normalize / convert the Gaussian (mixture) likelihoods.
rmg <options> ...
Speedup the Gaussian evaluation.
lmo <log_min_out>(0.0) [av_ngps](1.0)
Set the relative flooring value to avoid zero probs for non evaluated Gaussians. If a non-zero <av_ngps> (average number of Gaussians per state) is given, the LMO is automatically decreased by log2(av_ngps). If an HMM is given and <av_ngps> is zero, then <av_ngps> is derived from the HMM.

Author
Kris Demuynck
Date
05 May 2011