[DECtalk] DECTalk At High Speaking Rates
Don
Text_to_Speech at GMX.com
Tue Nov 1 02:39:18 EDT 2022
On 10/31/2022 9:57 PM, Jayson Smith wrote:
> They actually experimented with raising the sample rate to 22K from the default
> 11K. First, apparently this can only be done with the floating point vocal
> tract model and not the integer one, and I don't understand enough about the
> two models to know why. Also, there are some other problems I don't quite
> understand that prevent this from actually being a viable thing, at least in
> the short term. The experimental 22K support makes a few consonants have more
> highs, but otherwise sounds just about like the 11K version.
Of course; Nyquist requires > 2X the highest frequency to be reproduced.
I think you can get away with this on a PC as the soundcard hardware
is designed with an anti-aliasing filter set for ~20KHz audio. Making
the change on a hardware DECtalk would be less impressive as it,
no doubt, has its filter set for ~5KHz.
But, they didn't, also, change the parameter update rate, did they?
The original MITalk software just made step-wise changes to the
parameters at ~5ms intervals. If you are speaking faster, then
you are getting fewer -- coarser -- changes to the filters.
I've no idea how that would translate into intelligibility. OTOH,
if it would *improve* intelligibility, then you'd think they would
have made a corresponding change in the 200Hz update rate for
the "nominal" speaking rates to "improve" its intelligibility, no?
More information about the Dectalk
mailing list