On Mon, 27 Jun 2011, Alan W Black wrote:
message from Alan W Black <a...@cs.cmu.edu> to festival-talk = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = Heather Dewey-Hagborg wrote:Could anyone point me toward a brief explanation of how pitchmarks are used in diphone synthesis? I am continuing to tweak my festvox generated voiceand am trying to figure out how important the pitchmarks are ie. if I shouldtry to really tweak them to the point of editing by hand etc.The are pretty important. They are used to define join points and to define the pitch of the diphone (to know how much it has to be varied to the target).Extracting good pitch marks is hard. For some voices it works well and for others it does not. Four our system, the pitch marks must align with the largest peaks in the signal (in voiced regions). There a parameters that you can tune in bin/make_pm_wave, but I have to admit even I find them to be somewhat magic. The key ones are -min and -max though.There is some discussion here http://www.festvox.org/bsv/x862.html
You may want to try an alternative pitchmarker, program 'epochs' by David Talkin, which is part of ESPS, the Entropics Signal Processing System. In my experience it works much better than make_pm_wave / pitchmark without tweaking the parameters.Once I compared 4 different pitch trackers with their default parameters to manually corrected reference pitch tracks. ESPS get_f0, also by David Talkin, won by far. I never got round to do the same experiment with pitch markers, but I'd bet epochs would win. I know for sure it figures out the polarity of the signal (it tries both hypothesis - if you know the polarity, you can tell it and it runs twice as fast).
I have attached two wrapper scripts for epochs. ESPS itself used to be commercial software but is now available free here: http://ldc.upenn.edu/myl/esps60.6.linmac.src.tgz
Volker -- Volker Strom www.cstr.ed.ac.uk/ssi/people/vstrom.html The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.
pitchmark.tgz
Description: GNU Zip compressed data
_______________________________________________ Festlang-talk mailing list Festlang-talk@lists.berlios.de https://lists.berlios.de/mailman/listinfo/festlang-talk
