[hts-users:04010] Re: HMM parameters vs quality

[Matt Shannon <sms46@xxxxxxxxx>]

Hi Karthik,

I have previously investigated the change in TSLP (test set log 
probability) and MCD (mel cepstral distance) for different numbers of 
states and MDL clustering thresholds for the autoregressive HMM:

> http://mi.eng.cam.ac.uk/~sms46/papers/shannon2013autoregressive.pdf

but not for the standard HMM synthesis framework, I'm afraid.  I can say 
that for the standard HMM synthesis framework, using an MDL clustering 
threshold lower than the default value of 1.0 gives a better MCD score.

Thanks,

Matt


On 07/02/14 07:33, Sébastien Le Maguer wrote:
> Hi,
>
> I do not know any publications which summarize a complete evaluation of
> parameters used in HTS. Nevertheless, you can find pieces of what you
> want in the paper which are presenting the different concepts used in
> HTS, even if it's more for the recognition than the synthesis.
>
>
> Kind regards,
> Sébastien Le Maguer
>
>
> Karthik Krishnan writes:
>
> > Hi allI have been using the default configure parameters and am trying to experiment with varying options. Are there any references on the quality impact with various parameters ? I am hoping someone might have tried it and I would rather start with those since the training takes some time. Any pointers is much appreciated..Thanks<span style="background-color: transparent;">Karthik</span>