|
SPRAAK
|
All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Groups Pages
|
SPRAAK
|
one pass of viterbi training for HMM-models More...
Functions | |
| SprHmmHmm * | spr_vitpass_do (const SprCorSegOD *dbase, const SprHmmReadOD *acmod, const SprHmmSetupOD *eval, const SprLexOD *lex, const char *beam_info) |
one pass of viterbi training for HMM-models
spr_vitpass <-beam beam search info> [-ct FileName] <-c Corpus> [-range b:e](0:-1) [-ssp script](SPR_BSS_DEV_NULL) [-obs ObsDir] [-suffix FileSuffix](sam) [-arcd FileName] <-h FileName> [-g FileName] [-sel FileName] [-am_opt Options] [-top_n Number(s)](0) [-rmg rmg_params](no) [-LMout Value](-100) [-NOGS](flag: no gauss sel) <-ci unit description files> [-cd unit description files] <-d dictionary> [-unwind unwind format] [-ho FileName] [-go FileName] [-selo FileName] [-MinCount Value](1.0) [-LMsil Value](-10.0) [-LMspch Value](-10.0) [-REP](flag: relative minimal probabilities) [-LMtrans Value](-10.0) [-sf Factor](1.0) [-pool PoolingString](no) [-vm Value](0.0) (-u==-ci)
| -beam<em>beam | search info A beam search algorithm will be used. The extra parameters that can be specified are explained in the manual page of load_search_space() |
| -ct<em>FileName</em><a | name="spr_vitpass.ct" class="el"> Write out the accumulated statistics. If specified, and no dedicated output models (options -ho, -go, -selo) are given, only the statistics will be written out and no new acoustic model will be created/written. |
| -c<em>Corpus</em><a | name="spr_vitpass.c" class="el"> File with corpus entries or segmentations. |
| -range<em>b:e</em><a | name="spr_vitpass.range" class="el"> Optional begin and end entry the corpus/segmentation file. Counting starts at 0. |
| -ssp<em>script</em><a | name="spr_vitpass.ssp" class="el"> The signal processing script used to preprocess the input data. |
| -obs<em>ObsDir</em><a | name="spr_vitpass.obs" class="el"> Observation directory name. |
| -suffix<em>FileSuffix</em><a | name="spr_vitpass.suffix" class="el"> File suffix of the observation files (without leading '.'). |
| -arcd<em>FileName</em><a | name="spr_vitpass.arcd" class="el"> Unit file name (.arcd or .cd format). |
| -h<em>FileName</em><a | name="spr_vitpass.h" class="el"> The input HMM file. |
| -g<em>FileName</em><a | name="spr_vitpass.g" class="el"> The input MVG file (gaussians). |
| -sel<em>FileName</em><a | name="spr_vitpass.sel" class="el"> The input select file name (tied gaussian). |
| -am_opt<em>Options</em><a | name="spr_vitpass.am_opt" class="el"> Extra options for loading the acoustic model. A non-default acoustic model can be selected by having '=<am_type>;' as first option. See cwr_am_tbl.c for a list of acoustic models available. |
| -top_n<em>Number(s)</em><a | name="spr_vitpass.top_n" class="el"> Only take the top-N gaussians into account when calculating output probabilities. If one value is given, it is used for all mixtures. Else a value per mixture must be given, separated by commas. Use '0' to set top_n to the number of gaussians in the mixture. |
| -rmg<em>rmg_params</em><a | name="spr_vitpass.rmg" class="el"> The parameters for the quick selection of gaussians. If one value is given, it is used for all mixtures. Else a value per mixture must be given, separated by commas. Use 'no' if no quick selection is wanted. See rm_gauss.c for a description of the parameters. |
| -LMout<em>Value</em><a | name="spr_vitpass.LMout" class="el"> Floor the state likelihoods of an observation using a fraction of the unconditional likelihood of the observation (weighted sum of the state likelihoods). Practically necessary if only few gaussians are evalutated (-top_n or -rmg options). The value given offset an automatically determined log10(fraction). Use -100 to turn the flooring off, and 0.0 to use the default.automatically. |
| -NOGSflag | no gauss sel Forgo the (sentence level) lexicon based Gaussian selection. The lexicon based Gaussian selection speeds up the decoding but may interfere with score normalization techniques that assume all Gaussians were evaluated. |
| -ci<em>unit | description files The two units description files seperated by white-space. The first file just lists the units (phones). The second file describes the context dependencies. |
| -cd<em>unit | description files The two units description files seperated by white-space. The first file just lists the units (phones). The second file describes the context dependencies. |
| -d<em>dictionary</em><a | name="spr_vitpass.d" class="el"> Dictionary file name. |
| -unwind<em>unwind | format Define the parameters to modify the word transcriptions. See spr_cwr_lex_desc_read() for more details. |
| -ho<em>FileName</em><a | name="spr_vitpass.ho" class="el"> The output HMM file. Equals the input HMM if not specified (and no count file is requested). |
| -go<em>FileName</em><a | name="spr_vitpass.go" class="el"> The output MVG file. Equals the input MVG if not specified (and no count file is requested). |
| -selo<em>FileName</em><a | name="spr_vitpass.selo" class="el"> The output select file name (tied gaussian). |
| -MinCount<em>Value</em><a | name="spr_vitpass.MinCount" class="el"> Minimal count required for the state statistics to be updated. |
| -LMsil<em>Value</em><a | name="spr_vitpass.LMsil" class="el"> Minimal (logarithmic) probability for silence. |
| -LMspch<em>Value</em><a | name="spr_vitpass.LMspch" class="el"> Minimal (logarithmic) probability for speech. |
| -REPflag | relative minimal probabilities For each label probability / gaussian weight, a different minimal value is calculated from the value given with the option -LMspch, proportional to its relative occurrence in the data. For explanation, see ct_to_oprob2(). |
| -LMtrans<em>Value</em><a | name="spr_vitpass.LMtrans" class="el"> Minimal (logarithmic) probability for transition. |
| -sf<em>Factor</em><a | name="spr_vitpass.sf" class="el"> Smoothing factor: smooth the output HMM with the input HMM (mode=TRAIN), or with uniform probabilities (mode=INIT). |
| -pool<em>PoolingString</em><a | name="spr_vitpass.pool" class="el"> P: Pooled: ovlen dimensional sigma vector is the same for all gaussians. FP: Fully Pooled: only one sigma remains, it is the same for all gaussians AND dimensions. |
| -vm<em>Value</em><a | name="spr_vitpass.vm" class="el"> Lower limit on the sqrt(variance) relative w.r.t. the weighted average of the variances. |
Performs one pass of viterbi training for HMM-models. Reestimates the probabilities in a Hidden Markov Model in order to have a higher total probability for a given corpus file. The restimation (a derivate of the Expectation Maximization algoritme) is based on the optimal path only (see fbpass for a more general but slower reestimation). The newly estimated probabilities can be smoothed with the old probabiltities. Also hard minimal values for different probabilities in the HMM can be set (options: LMsil LMspch LMtrans). The minimal observation probability (option LMout) has only an effect on the alignation faze (finding the optimal path), and is no hard limit on the probabilities in the HMM.
For a normal training the observation distributions in each state are modified in order to have a maximal total probability for all observations assigned to that state during the viterbi alignement. In the discrete case this results to setting the probability of each label proportional to the number of times the label was observed in that state (label count).
1.8.6