[DECtalk] Some DECtalk history and what I think we can and can't reasonably do

[Alex H.]

Well, to us,, we never really heard later versions of DT, only the 
classics from the 90's, so forgive us if we compare the new attempts to 
prior versions - it's not like we have a huge library of source code to 
just browse at will and endless samples of every version.... so... yeah.

Wanna know what's been wrong with the samples and attempts posted to 
this list a few months ago for the sapi dectalk? I'll tell you.

The voices were clipping and squawking, and all the voices sounded like 
they had a speech problem. Perfect Paul wasn't perfect as most of us 
have heard before. The voices themselves sound not like DECTalk at all, 
they also drop out in volume, just like a human cuz it's using HLsyn to 
make it sound more natural.
I've heard DT 4.2cd, 4.3, 4.4, 4.61, 4.62 and 4.64. But since you've 
pointed out before that version numbers don't matter to speak, is this 
even important anyway or are we just listening to the same code with 
minor tweaks to get the various versions we know?

Disable HLsyn in the new product, and it'll suck less. I like forment 
based synths, not ones that try and sound human, because I and others 
are used to classic forment non-HLsyn versions of DECTalk. True that 
HLsyn is still formant but it's trying to sound real and have human 
articulation, and knowing that I can understand why this version sounds 
different. It's just not what we're used to, that's all. Some Joe Blow 
off the street who has never heard synthesized speech can't understand 
Eloquence from DECTalk from Espeak anyways, so this point of 
understanding speech is a moot one.  They'd be better off using Cepstral 
or some human-sampled synths and wasting their hard drive space. This is 
being targeted at a relatively small group of people who have used 
DECTalk before and like it, so i think we're safe there. I'd consider 
giving HLsyn another shot if it was completed. But as always, corporate 
America screws everyone over in the end, and that was the case with 
Dectalk. So much so, that Fonix wanted to make FonixTalk and specificly 
try and make it sound human. The result sucks.


Alex
On 8/3/2011 11:17 AM, ebruckert Bruckert wrote:
>    First of all let me make you aware that I use DragonDictate, as I 
> can't see very well and proofreading is quite painful so you'll have 
> to forgive and interpret from mistakes the DragonDictate may make. It
>    I was taught about form and speech synthesis by Dennis Klatt, and 
> by reading but before my involvement with him I knew next to nothing. 
> One of the questions in the early days was could you achieve higher 
> intelligibility by super articulation and do better than natural 
> speech. What testing revealed was really two things. At normal 
> speaking rates the answer always seem to be that the closer you 
> matched to real speech the better the intelligibility at higher 
> speaking rates above that which humans could normally achieve things 
> were little different and I'm not going to go into the specifics of 
> what we did to make things better at high speed other than to say they 
> were based on knowledge of speech perception.
>      The second thing we learned is that listening to a synthesizer 
> has a very fast but steep learning curve. Somewhat analogous to 
> learning to understand a person with a strong dialect or speech 
> impediment. One of the problems we encountered is that people often 
> preferred the version they were used to over any succeeding version. 
> But actual tests did not support the preference.
>      One example is the way tilt was done inside DECtalk. The original 
> mechanism was a crude approximation of spectral tilt. Dennis before he 
> died developed a much more accurate (meaning matching human 
> production) tilt filter that was not able to be incorporated to a 
> later date. As a point of interest Dennis was so dedicated that he 
> last modified the DECtalk code 3 days before he passed away. So the 
> spectral tilt was changed and this changed what you might consider the 
> tone control on an old radio or record player. That is just one of 
> many reasons why DECtalk change slightly over the years.
>       The 5.0 DECtalk Incorporated the work of Prof. Ken Stevens who 
> was Dennis is blessed MIT and close friend. The 5.0 code unfortunately 
> did not yield the expected results, but we did learn a lot from the 
> attempt. This
>        there are even some changes to DECtalk that would change the 
> way it sounds from any particular version, such as Intonation that I 
> am unwilling to revert because I know for a fact that they caused loss 
> of information. So my goal is very simple I am working to create a 
> very functional intelligible DECtalk to put back out, I am unwilling 
> to try and make it sound exactly like any given person wants to. I 
> have been through this before and the year is very sensitive and if 
> you directly comparing two versions side-by-side you not testing 
> anything but whether did the same and that is an exercise in futility. T
> Any specific issues I can address. Secondly as a word of warning to 
> listeners providing feedback. The other thing we've learned is that 
> listeners are excellent at deciding that something is not right, but 
> are absolutely terrible at exactly pinpointing the problem. The reason 
> for this is quite simple people judge the output as speech which it 
> only kinda is, by this I mean that a synthesizer can make mistakes 
> that humans cannot possibly do and as a consequence can't possibly 
> recognize. An example of this is that after so many years of working 
> with it I have learned to hear a foreman that's moving too rapidly, 
> but most people cannot hear it. This is because to make life easy we 
> try to lead nor stuff that's not important in our language, such as 
> the nasal lifestyles in French or the retro flex ours in American 
> English which is Sheehan have a heckuva time hearing.
>
>
> _______________________________________________
> DECtalk mailing list
> DECtalk at bluegrasspals.com
> http://bluegrasspals.com/mailman/listinfo/dectalk


-- 
Sent via Thunderbird.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://bluegrasspals.com/pipermail/dectalk/attachments/20110803/8ad01c17/attachment.html>