Sacha Krstulovic wrote:
In the HTS-SLT demo, during the HHEd context-clustering phases (CXCL1 and CXCL2), the .hed files for cmp and lf0 call the clustering commands according to the model's state order (i.e., for 3 states models, first operate the clustering for state 2, then for state 3, then for state 4 across the models). In the way HHEd operates, does the overall result of the clustering depend on the order of these commands?
A separate tree is built for each state / stream combination. So, one way to make this step faster is to build these trees in parallel with separate runs of HHEd, then merge the resulting MMFs (there is no HTK tool to do MMF merging though - you need a script)
can these clustering operations be safely considered orthogonal from the point of view of MDL optimization?
Yes, they are independent. Simon -- The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.