[SRILM User List] Confusion network combination

[Andreas Stolcke]

On 10/1/2013 5:09 AM, Cristina Guerrero wrote:
> Hello,
> I am looking for information to accomplish 
> confusion-network-combination with the SRILM toolkit. I want to use 
> the lattices generated by different speech recognizers over the same 
> speech segment. I haven't found a detailed description of the steps to 
> follow, so here is what I'm doing right now:
> 1- Extract lattices from the various recognizers (With HTK in my case)
> 2- Take one of these lattices as a starting point and convert it into 
> a confusion network MESH0 (lattice-tool -read-htk -in-lattice LATTICE0 
> write-mesh MESH0). *I know -posterior-prune can be applied to the 
> lattice before building the mesh for better results according to the 
> "Finding consensus.." paper.
> 3- Then, take the next lattice (LATTICE1) and merge it with the 
> previously generated mesh ( lattice-tool -in-lattice LATTICE1 
> -init-mesh MESH0 -write-mesh MESH1).
> 4- And repeat the merging step (3) using the previous mesh to 
> initialize the next lattice.

What you're doing works, but it is a roundabout way to perform confusion 
network combination and you don't have control over the weighting of the 
posterior probabilities in each input CN.  A more straightforward 
approach is to dump the various CNs for each utterance, then combine 
them in one step using the command

nbest-lattice -use-mesh -lattice-files FILE

where FILE contains a list of CN files and associated weights (see man 
page for details).

Andreas