NAME

hmm_nnembed - Embedded re-estimation of neural network targets.

AVAILABILITY

cslush/pkgs/hmm

SYNOPSIS

 hmm nnembed ?-prune float? ?-minmodel float? hmmlistob nnprobob fbvar

PARAMETERS

hmmlistob

A model-list object which contains pointers to each individual HMM model. The model-list object is created by the concatenation of individual HMM model pointers. (see hmm_concat(n).

nnprobob

Probability vector computed from the neural-network forward pass.

fbvar

A Tcl variable used to store all temporary memory needed for HMM embedded re-estimation.

OPTIONS

-prune float [Default = 300.0]

Embedded model re-estimation pruning threshold.

-minmodel float [Default = 10.0]

Embedded model re-estimation minimum model probability.

DESCRIPTION

hmm nnembed reestimate neural-network targets using the forward/backward algorithm. This technique is used to create a set of soft targets to train towards, rather than the conventional 1/0 targets.

Each successive call to hmm nnembed will return a neural-network target object (nntargetob) which can be used in conjunction with the data selection (see framepick_eval(n)) to generate a neural network training file. The parameters are then updated off-line using the fbtrain executable.

The options -prune and -minmodel are used to control the number of models which are kept active during the training. If an over pruning error is caused, due to too tight pruning, the pruning threshold is automatically relaxed and the forward/backward algorithm is restarted for that particular utterance. The pruning then is set back to its original setting.

RETURNS

hmm nnembed returns a neural-network target object which can then be written to disk using hmm nnwrite, for off-line parameter updating.

SEE ALSO

hmm_concat(n), hmm_infolist(n), framepick_eval(n), hmm_nnwrite(n), fbtrain(1)