[hts-users:00949] Re: Questions about certain concepts

[Simon King <Simon.King@xxxxxxxx>]

marc sobhy wrote
> 1. i can't imagine how the calculation of global means and variance 
> can be done.more explanation plz

It's quite simply the mean and variance of all the observation vectors - 
no labels or models are involved.

>  
> 2. how the HERest can determine the phoneme boundaries (this automatic 
> segmentation ) which algorithm can do this and how it do?

You are thinking about EM in the wrong way - there is no explicit 
segmentation (i.e. a single fixed alignment between labels or models, 
and data) in HERest - *all* possible alignments of models and data are 
considered. All that is needed is the *sequence* of models required for 
each training utterance. The *alignment* of these models to the data 
does not need to be specified.

In HERest, each possible alignment between model states and the data is 
weighted by its likelihood given the current model.

In the first iteration or two of HERest, poor alignments (between model 
and data) will be considered to be just as likely as good alignments. 
Then, as the model parameters become better trained, the good alignments 
will become more and more likely and the poor alignments will become 
less and less likely.

Simon