[hts-users:02962] Re: element of WUW negative in HMgeTool

[Ninh Khanh Duy <ninhkhanhduy@xxxxxxxxx>]

Hi,

It seems that the 'negative WUW element' error happens when some variance component become negative after MGE training. Could you check that?

Regards,

Ninh K.D.

2011/7/19 那兴宇 <nxy-yzqs@xxxxxxx>

Hi,

 

Yes, I encountered the same problem as yours. But I haven't checked the code yet.

Now I re-tested "mgc+lf0+bap" streams again to see if there is anything wrong with my script.

 

Na Xingyu

At 2011-07-19 15:28:45，"Ninh Khanh Duy" <ninhkhanhduy@xxxxxxxxx> wrote:

Hi,

I'm not sure about that. So far I tested MGE reestimation on mgc stream only and it worked fine except that no state boundary refinement occurred as expected.

Regards,

Ninh K.D.

2011/7/19 那兴宇 <nxy-yzqs@xxxxxxx>

Hi

 

When I switched on -g option, the ERROR even occured several samples earlier.
I tested the single MGE estimation command earlier, and it worked fine for "mgc+lf0+bap". Now I remove the 'bap' stream, with the same database, do you think this is relevant?

 

Na Xingyu

 

At 2011-07-19 12:12:40，"Ninh Khanh Duy" <ninhkhanhduy@xxxxxxxxx> wrote:

Hi,

Could you try switching on -g option?

Ninh K.D.

2011/7/19 那兴宇 <nxy-yzqs@xxxxxxx>

Hi,

 

Sorry for the wrong discription about the error. It occured whatever I set the iteration counters at. What I did is like this:

I added HSMMAlign and HMgeTool in the demo scripts twice. One is after the re-estimation of the first clustered model (ie, clusterd.mmf and full.list) and the other is after the re-estimation of the second clustered model (ie, re_clustered and full.list), both in cmp model dir. After the error occured, I deleted the first MGE step, so there is only one MGE re-estimation after all the ML trainings, with '-i 1 5'. But the 'negative WUW element' error still came out. The model is what I used for synthesis and the engine format of it worked fine. So I was confused.

Anyone give a hint, please? Thanks a lot!

 

Na Xingyu
 

At 2011-07-18 09:45:01，"那兴宇" <nxy-yzqs@xxxxxxx> wrote:

Hi,

I encountered a new problem. When I do MGE training, an error occurred at iteration 1:

ERROR [+9999] Cholesky_Factorization: (253,253)-th element of W'*U(-1)*W is negative.

The iteration bound of HMgeTool  is set to be "-i 0 5". When I used "-i 0 5", This error didn't occur.

And another question is, why the cmps are processed orderedless when the iteration number is larger then 0?

Thanks in advance,

Na Xingyu