[SRILM User List] Posterior Probability : HTK

[Andreas Stolcke]

The first step is to compute posterior probabilities for arcs and nodes 
in your lattice, using the forward-backward algorithm.
The posterior probability is the sum of the scores of all paths going 
through an arc/node, normalized by the sum over all paths through the 
lattice.
This is implemented by the lattice-tool -write-posteriors option (the 
output format is different from HTK format though).
It is important to scale the combined acoustic/language model scores, 
check the -posterior-scale option.

Often one wants posterior probabilities at the word level, and combine 
all word hypotheses that occur at the same "position" in the lattice.
For this you can build a word confusion network or "word mesh".  This is 
done by the lattice-tool -write-mesh option.

For an introduction to these concepts you might want to check the 
article 
http://www.speech.sri.com/cgi-bin/run-distill?ftp:papers/CSL2000-consensus.ps.gz, 
but note that the confusion network algorithm in SRILM is not the same 
as described in there.

Andreas

On 3/8/2012 10:39 PM, Vinay Shashidhar wrote:
> Hi Guys,
>
> I have a read a lot of papers regarding posterior probability being a
> more robust and speaker independent features, but how does one
> calculate it?
>
> I am using HTK and am doing forced alignment. All i get is the
> likelihood scores.
>
> Thanks. Looking forward for your help.!
>
> regards
> Vinay
> _______________________________________________
> SRILM-User site list
> SRILM-User at speech.sri.com
> http://www.speech.sri.com/mailman/listinfo/srilm-user