[SRILM User List] Does sausage-making use timing information?

Andreas Stolcke stolcke at speech.sri.com
Sat Mar 13 15:26:30 PST 2010

On 3/13/2010 4:09 AM, Tim Kempton wrote:
> Hi
> Does lattice-tool use timing information (if available) when transforming
> lattices to word meshes (aka sausages/confusion networks)?
> > From the manual pages it looks like the recommended reading for meshes is
> the following paper: L. Mangu, E. Brill, and A. Stolcke, ``Finding Consensus
> Among Words: Lattice-based Word Error Minimization.'' Proc. Eurospeech, vol.
> 1, 495-498, 1999.
> The section on intra-word clustering implies that timing information can be
> used. However some experiments with lattice-tool indicates that timing
> information is possibly ignored e.g. one mesh derived from a lattice with
> timing information (HTK) was very similar to a mesh derived from the same
> lattice without timing information (PFSG). Although there may be other
> reasons for this.
lattice-tool does in fact use a very different (and more efficient) 
algorithm for sausage construction, and uses no time information,
only the topology of the lattice. However, the time stamps from the 
input lattice can be carried over to the output sausage if the 
-acoustic-mesh option is used.
> The reason I'm asking all this, is because I'd like to combine results from
> multiple recognizers. I like the nbest-rover approach but I'd like to make
> use of timing information too, partly because the output of the recognizers
> are very different. So I'm hoping I can combine multiple recognizer lattices
> (with -operation or), and then transform this to a mesh.
I guess I don't quite see how using times helps you in this case.  But 
the procedure you suggest should work in principle.


More information about the SRILM-User mailing list