|
SPRAAK
|
All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Groups Pages
|
SPRAAK
|
Smooth the mixture weights in HMM's retrained on little data. More...
Smooth the mixture weights in HMM's retrained on little data.
spr_hmmsmooth <-i hmm_file> [-gi mvg_file] [-seli sel_file] [-u unit_fname] [-am_opt acmod_opt] [-o hmm_file] [-go hmm_file] [-selo sel_file] [-nu unit_fname] [-i2 generic_hmm] [-wi weight_em0](1.0) [-wo eqn_weight_orig](1) [-we eqn_weight_em](1) [-wg eqn_weight_generic](1) [-Nem NrIter](1) [-alpha em_smooth](1.0)
| -i<em>hmm_file</em><a | name="spr_hmmsmooth.i" class="el"> HMM-file to read. |
| -gi<em>mvg_file</em><a | name="spr_hmmsmooth.gi" class="el"> MVG-file to read. If specified, this MVG file is read instead of the MVG file given in the header of the input HMM. |
| -seli<em>sel_file</em><a | name="spr_hmmsmooth.seli" class="el"> SEL-file to read. If specified, this SELECT file is read instead of the SELECT file given in the header of the input HMM. |
| -u<em>unit_fname</em><a | name="spr_hmmsmooth.u" class="el"> UNIT-file to read. If specified, this unit file is used to read the input HMM instead of the file corresponding to the filename in the header of the input HMM. |
| -am_opt<em>acmod_opt</em><a | name="spr_hmmsmooth.am_opt" class="el"> Extra options for loading the acoustic model. |
| -o<em>hmm_file</em><a | name="spr_hmmsmooth.o" class="el"> HMM-file to write to. If not specified, the input HMM will be overwritten. |
| -go<em>hmm_file</em><a | name="spr_hmmsmooth.go" class="el"> MVG-file to write to. By default the input mvg file will be overwritten. |
| -selo<em>sel_file</em><a | name="spr_hmmsmooth.selo" class="el"> SEL-file to write to. By default is input sel file will be overwritten. |
| -nu<em>unit_fname</em><a | name="spr_hmmsmooth.nu" class="el"> Specify a new unit file to write the units to (or as input for rearanging the existing units). |
| -i2<em>generic_hmm</em><a | name="spr_hmmsmooth.i2" class="el"> An HMM-file containing more robust mixture weights. If not specified, the uniform distribution is used. |
| -wi<em>weight_em0</em><a | name="spr_hmmsmooth.wi" class="el"> The initial weights for the simulated EM iterations are obtained as a linear combination of the generic weights (uniform or drawn from the generic hmm if given) and the weights based on the observed (limited) data. The default value (1.0) means starting from the generic weights. |
| -wo<em>eqn_weight_orig</em><a | name="spr_hmmsmooth.wo" class="el"> Equation (in function of x, the number of data points observed for that state) to calculate the weight given to the original (possibly ill-estimated) weights. |
| -we<em>eqn_weight_em</em><a | name="spr_hmmsmooth.we" class="el"> Equation (in function of x, the number of data points observed for that state) to calculate the weight given to the weights obtained with simulate EM-training. |
| -wg<em>eqn_weight_generic</em><a | name="spr_hmmsmooth.wg" class="el"> Equation (in function of x, the number of data points observed for that state) to calculate the weight given to the generic distribution (see option -i2). |
| -Nem<em>NrIter</em><a | name="spr_hmmsmooth.Nem" class="el"> Number of simulated EM iterations. Use a higher value if no generic HMM is available. |
| -alpha<em>em_smooth</em><a | name="spr_hmmsmooth.alpha" class="el"> Smooth the simulated EM re-estimation by taking raising the likelihoods to the power alpha. |
Increase the level of tying in a reduced SC_HMM. This program assumes a well trained gaussian set combined will ill estimated mixture weights due to limited amounts of data (e.g. speaker dependent models). The mixture weights are smoothed using two methods:
The weights assigned to the original (possibly ill-estimated) weights, the EM-smoothed weights and the generic weights is specified using an equation, and hence the amount of smoothing can made dependent on the amount of data observed in each state.
1.8.6