[hts-users:03766] Re: small bug in HHEd

[Kei Hashimoto <bonanza@xxxxxxxxxxxxxxx>]

Hi Matt,

Thank you for your report.
We will fix the bug in the next version of HTS.

Thanks,

Kei


(2013/05/04 1:14), Matt Shannon wrote:
> Hi,
> 
> There is a small bug in clustering which results in slightly incorrect behaviour when variance flooring is used.
> 
> In the AccSumProb function in HHEd.c the line:
> 
>>     x+=-0.5*occ*(1.0+log(TPI*variance));
> 
> computes the auxiliary function value given a re-estimated parameter variance.  However this expression is only correct in the case of no variance flooring.
> 
> I have confirmed that this leads to a different number of leaves being chosen in some cases (for the autoregressive HMM, but the same should be true for the standard HMM).
> 
> Thanks,
> 
> Matt
> 
>