[SRILM User List] Hypothesis from a mesh

[Andreas Stolcke]

On 10/16/2013 6:35 AM, Cristina Guerrero wrote:
> I'm working with confusion networks/sausages.
> Observing the posteriors in each 'align' it seems that the command: 
> lattice-tool -read-mesh -in-lattice MY.MESH -viterbi-decode
> extracts the hypothesis with the lowest WER (it is with the highest 
> posteriors per align). Is this correct? From my observation, using 
> "-posterior-decode" doesn't extract the best-hypothesis out of a mesh.

Posterior decoding will extract the words with the highest posterior 
probability estimates for each alignment position.   Of course these are 
ESTIMATES, and even if you had accurate posterior probabilities the 
actual words spoken could be different.   That's the nature of a 
probabilistic classifier!

Andreas