[hts-users:02804] Re: average log prob per frame higher than zero...

[Hui LIANG <tshlmail-hts@xxxxxxxxx>]

Hi Joelle,

I don't think those numbers (i.e. 'average log prob') are probabilities. They are actually probability density values, because they were derived from *continuous* Gaussian functions.

Correct me if my understanding is wrong.

Thanks,
Hui LIANG


--- On Wed, 20/4/11, Joelle Tilmanne <joelle.tilmanne@xxxxxxxxxxx> wrote:

> From: Joelle Tilmanne <joelle.tilmanne@xxxxxxxxxxx>
> Subject: [hts-users:02796] average log prob per frame higher than zero...
> To: hts-users@xxxxxxxxxxxxxxx
> Date: Wednesday, 20 April, 2011, 17:45
> 
> 
> Dear HTS-users,
> 
> The "average log prob per frame" is often used as an
> indicator of how well the HMM matches the training sequence.
> But I did not found an explanation on how that number was
> calculated and I can not explain why a logarithmic
> probability can have values higher than 0 (or why a
> probability can be higher than 1). In my case and in the
> literature, I find average log prob per frame = several
> hundreds and more ...
> 
> Does somebody have an explanation?
> 
> Thank you very much for your help,
> 
> Joelle