Decision tree based splitting of gaussians for HMMs with tied gaussians. More...

Detailed Description

Decision tree based splitting of gaussians for HMMs with tied gaussians.

spr_tree_gauss_split [-global](flag: ) <-hi HMMFileName> <-ho HMMFileName> [-u UnitFileName]
    [-gi MVGFileName] <-go MVGFileName> [-seli SelFileName] <-selo SelFileName> <-ci CIphones>
    <-cd CDphones> <-q Questions> <-tree Trees> [-nfr_per_gauss Minimal_nfr_per_gaussian](128)
    [-balanced](flag: ) [-max_overlap Max_overlap](0.0)

Parameters

-globalflag	The tree is a global tree, which means that phones that have no example at all should not be purged.
-hi<em>HMMFileName</em><a	name="spr_tree_gauss_split.hi" class="el"> Input HMM.
-ho<em>HMMFileName</em><a	name="spr_tree_gauss_split.ho" class="el"> Output HMM.
-u<em>UnitFileName</em><a	name="spr_tree_gauss_split.u" class="el"> Input Unit file name.
-gi<em>MVGFileName</em><a	name="spr_tree_gauss_split.gi" class="el"> MVG file name.
-go<em>MVGFileName</em><a	name="spr_tree_gauss_split.go" class="el"> Output MVG file name, with split gaussians.
-seli<em>SelFileName</em><a	name="spr_tree_gauss_split.seli" class="el"> Input SELECT file name.
-selo<em>SelFileName</em><a	name="spr_tree_gauss_split.selo" class="el"> Output SELECT file name.
-ci<em>CIphones</em><a	name="spr_tree_gauss_split.ci" class="el"> File with context independent phones.
-cd<em>CDphones</em><a	name="spr_tree_gauss_split.cd" class="el"> Input File with context dependent phones.
-q<em>Questions</em><a	name="spr_tree_gauss_split.q" class="el"> File with the binary questions.
-tree<em>Trees</em><a	name="spr_tree_gauss_split.tree" class="el"> File to read the tree(s) from on which the HMM is based.
-nfr_per_gauss<em>Minimal_nfr_per_gaussian</em><a	name="spr_tree_gauss_split.nfr_per_gauss" class="el"> Minimal number of frames in a gaussian: a tree node will not be chosen for splitting if this results in a gaussian with occupancy lower than this minimal value.
-balancedflag	That node in the tree is chosen that results in two gaussians with as much as possible balanced occupancies.
-max_overlap<em>Max_overlap</em><a	name="spr_tree_gauss_split.max_overlap" class="el"> For tree pruning: see tree2model.

Decision tree based splitting of gaussians for HMMs with tied gaussians. Splits each gaussian in decision tree based context dependent acoustic models in two new gaussians. Decides, based on the decision tree, which states (in the HMM system with tied gaussians) will use the one gaussian, and which states will use the other. Both gaussians will be copies of the original gaussian. They should be differentiated through further training. For more information, see the author's PhD thesis, section 5.3.6.

Author: Jacques Duchateau

Date: April 1998

Revision History:: JD, March 1999: added to package HMM version 70