Hi YangWang:
Thank you for such a detailed answer. I guess I got the answer, Thanks again sincerely! > Date: Thu, 7 Jun 2012 10:56:06 +0800 > From: yangwang84@xxxxxxxxx > Subject: [hts-users:03336] Re: [hts-users:03334] RE: [hts-users:03326] Re: [hts-users:03323] Questions of the untied stage in the sd training.pl? > To: hts-users@xxxxxxxxxxxxxxx > > Hi, ArthurLeo, > > It is stated in HTK book that "The philosophy of system > construction in HTK is that HMMs should be refined incrementally." It > is also true in HTS. > For the first time of clustering, we get full context HMMs which is > simply cloned from mono phone and re-estimate for just once, so the > precision of initial full context HMMs is generally acceptable but not > very high. When you cluster these initial full context HMMs, the > robustness of HMMs is much improved and thus could be re-estimated for > $nIte times rather than just once. When we untie and re-estimate these > clustered full context HMMs, we get higher precision of these untied > full context HMMs in principle. Then the second clustering operation > is needed to cluster these untied full context HMMs which is refined > compared with the cloned and re-estimated intial full context HMMs. > In summary, the cluster-untie-recluster operation sequence is > better than clustering only once to get higher precision and > robustness of HMMs in principle. > > I did not try to cluster-untie-recluster-reuntie-recluster2, as I > guess its improvement will not be worthy when considering its > increased training time. > > YangWang > > > 2012/6/6 ArthurLeo <bin007.zhao@xxxxxxxxxxx>: > > thank you for your reply!but I still have the question why doing fu rther > > untie, re-estimate, recluster and herest > > HMMs can get the better re-estimation accuracy? Have the inevitable reasons > > in theory? or can you explain in theory simply? > >> Date: Wed, 6 Jun 2012 09:31:50 +0800 > >> From: yangwang84@xxxxxxxxx > >> Subject: [hts-users:03326] Re: [hts-users:03323] Questions of the untied > >> stage in the sd training.pl? > >> To: hts-users@xxxxxxxxxxxxxxx > > > >> > >> Hi, ArthurLeo, > >> > >> You asked about "{cluster, herest}, {untie, herest}, {recluster, > >> herest}". When we further untie, re-estimate, recluster and herest > >> HMMs, we usually get better re-estimation accuracy than just cluster, > >> herest. You may verify that the corresponding HMM definition files are > >> a little different. Through my informal listening test, I found the > >> synthesi zed speech is indeed a little better. > > > >> > >> YangWang > >> > >> > >> 2012/6/5 ArthurLeo <bin007.zhao@xxxxxxxxxxx>: > >> > Hi all, > >> > When I read the SD training.pl(hts2.2 version),I found that there is a > >> > untied stage, I want to know why do this stage? why not use the before > >> > stage > >> > result cluster.mmf(after the herest) as the final model? why do the > >> > untied > >> > stage then clustered, next using herest result as the final model? what > >> > is > >> > the function of > >> > adding the untied stage? what is the benefit? Can someone give me an > >> > answer? Thank you in advance! > >> > |