[Kde-accessibility] Factors that affect accuracy

[Peter Grasch]

On Tue, 2012-10-02 at 10:31 -0300, Steve Cookson wrote:
> >But the most important advice I can give you for the pilot phase is to
> >activate sample harvesting in Simond:
> 
> What an excellent idea.  I don't know why we didn't think of it before! 
> 
> So we can have a selection of machines out there (say 5) with sampling.  The
> new samples will not, I guess be included in the local models, but at the
> end of the pilot we can download the files and included into the central
> model using the training module? Is that how it works?
Yes.


> Would there be any disadvantage in having sampling on permanently active on
> all machines, then if a user has difficulty, we can download his own
> personal corpus and include it in the model?
No other disadvantage than that Simond does not enforce any kind of
quota and that the audio samples are, although segmented, not
compressed.
You'll eventually run out of disk space if you don't set up e.g. a
cronjob to remove old samples once in a while. 

But yeah, keeping the samples around can help tremendously with problem
cases - either by adapting the model for those users who are having
trouble with recognition accuracy or simply for easier diagnostics.

Best regards,
Peter