[SRILM User List] Some questions about lattice-tool

[Andreas Stolcke]

On 3/21/2010 8:48 PM, wei chen wrote:
> Hi all,
> I am transforming lattices generated by speech recognition to 
> confusion networks using lattice-tool, and  because confusion network 
> decoding is based on minimum word error rate(WER), it would reduce 
> WER, but in my experiments, the performance became worse after 
> confusion network decoding, so I  am confused if there exist some 
> errors in my command line as below,
>
> [step 1] : lattice prunning
> bin/lattice-tool.exe  -read-htk -in-lattice-list [lattice_list] 
>  -write-htk -out-lattice-dir [lattice_output_dir]  -overwrite 
> -posterior-prune 0.001
>
> [step 2]:  transformaing lattice to confusion network
> bin/lattice-tool.exe  -read-htk -in-lattice [lattice]  -acoustic-mesh 
> -write-mesh [mesh_name] -htk-lmscale $gram_factor -htk-acscale 
> $HTKasscale -htk-wdpenalty $insert_prob  -overwrite -ignore-vocab 
> [ignore_dict]
>
You are doing several things to the lattices besides building confusion 
networks.
- pruning
- lmscale change
- acscale change
- wdpenalty change

First make sure all these measures don't change your results for the 
worse.  You should try doing 1-best viterbi decoding on the resulting 
lattices, and compare those results to CN decoding.

Andreas