[Subject Prev][Subject Next][Thread Prev][Thread Next][Date Index][Thread Index]

[hts-users:01141] Re: HTS2.1beta-HVite


Heiga ZEN (Byung Ha CHUN) wrote:

You can use HVite of HTS-2.1beta with *HMMs*.

and it is straightforward to convert your HSMMs to HMMs. Here is my recipe:

- in your HSMM, there is a dummy transition matrix which is probably shared by all models (this is what the HTS demo scripts do), so untie it, then retie so it is shared between models for the same centre phone (for example). Then, modify the transition matrices so they are the usual band-diagonal form (self transition and forward transition for all emitting states). Here is my .hed file for doing that using HHEd (models with 5 emitting states):

TR 2
UT {*.transP}
TI T_aa {(*-aa+*).transP}
TI T_ae {(*-ae+*).transP}
TI T_ah {(*-ah+*).transP}
[... etc for all phones]
AT 2 2 0.9 {*.transP}
AT 2 3 0.1 {*.transP}
AT 3 3 0.9 {*.transP}
AT 3 4 0.1 {*.transP}
AT 4 4 0.9 {*.transP}
AT 4 5 0.1 {*.transP}
AT 5 5 0.9 {*.transP}
AT 5 6 0.1 {*.transP}
AT 6 6 0.9 {*.transP}
AT 6 7 0.1 {*.transP}



- next, train the transition matrices - do one or two rounds of EM. You could tell HERest to only update the transition matrices (-u t) and leave the Gaussians unchanged, if you want your HMMs to be as similar as possible to your HSMMs


Simon


Follow-Ups
[hts-users:01142] Re: HTS2.1beta-HVite, Sacha Krstulovic
References
[hts-users:01135] HTS2.1beta-HVite, zhizheng wu
[hts-users:01136] Re: HTS2.1beta-HVite, Heiga ZEN (Byung Ha CHUN)
[hts-users:01137] Re: HTS2.1beta-HVite, zhizheng wu
[hts-users:01138] Re: HTS2.1beta-HVite, Heiga ZEN (Byung Ha CHUN)
[hts-users:01140] Re: HTS2.1beta-HVite, Heiga ZEN (Byung Ha CHUN)