Hi,

I'm using HTK for handwriting recognition (I produce my own 'mfcc' files, then the steps are identical). I succeeded going through all the first steps of the tutorial. I am now trying to go through writer (speaker..) adaptation. I do supervised adaptation.

The forced alignment is done with HVite, then I adapt with a 32-regression tree and a MLLR transformation.

I wish to do antoher step. So I force align again with HVite, given the MLLR transformation. The output file gives (obviously) the same phones (as it is supervised adaptation), but the time alignments change.

However, when I re-run HERest for MLLR adaptation with this new alignment, the output transformation is exactly the same as the precedent.

How is this possible? Do HERest no take into account the time alignments computed by HVite? How could I refine my transformation?

Thanks for your help! Don't hesitate to ask me command lines, or files, if you need it.

Anne-Laure

--- (Edited on 12/22/2009 8:06 am [GMT-0600] by biannelor) ---