> It will probably take a while before usable OS dictation software will be a realistticobjective

Hi Robin

To be honest in my opinion dictation software is a realistic goal right now. It requires some effort, but only a little one. With properly implemented adaptation you can reach reasonable accuracy in a short period of time. The rest is pure programmer work on handling corrections, input processing and so on.

This is stated to encourage other developers to join this effort actually, not to discourage anyone.

--- (Edited on 11/7/2009 03:57 [GMT+0300] by nsh) ---

Thanks Nickolay, perhaps I sounded to pessimistic.

I also think that we're relatively close, but it will take a while before we have dictation with 99% accuracy and incredible ease of use, right? I'd like you to prove me wrong before the end of the year, but believe it or not, you also need sleep occasionally.

I thought that a simpler short term goal could bring us faster to the ultimate goal, because if people see an app that works flawless, they are more willing to contribute speech or feedback. If they see an app that out of the box only offers 95% accuracy, they're more likely to wait for another year before even giving it a try.

Of course, there are exceptions to that rule, but we should not expect too much from the average person. Anyway, any good dictation solution also has spelling functionality, so it could be just one component that can be reused in the final app, and of course, the radiotelephony spelling could be just one possibility. Natural spelling would be nicer, but probably also a bit harder.

Slightly off topic, would it be helpful to have speech recordings of continuous spelling? Like v o x f o r g e

--- (Edited on 11/7/2009 1:16 pm [GMT-0600] by Robin) ---

Greetings all,

Here is some work I offer to voxforge

defining acrophonic cyclic alphabet input.

I have not had time to do much testing

of it's reliability or pratical usefulness.

http://g2p4j.sourceforge.net/ACA_alphabet.html

Best Wishes

paradocs

--- (Edited on 11/8/2009 1:00 am [GMT-0600] by paradocs) ---

Hi paradocs,

That's a very original idea. I think it's main advantage lies in the fact that people would not have to learn a new spelling alphabet, since they probably are very familiar with the alphabet of their language.

I think there are also some disadvantages.
The main one being that people would have to learn its concept, so there is still some extra effort required. Some might prefer learning a concept, others might prefer a list of words to memorize.

Another might be that some combinations are phonetically more closely related than the option in existing spelling alphabets that have stood the test of time, as you notice yourself. In particular the international radiotelephony alphabet was designed to be useful to communicate arbitrary sequences of letters, or at least it's used for than on a daily basis, so the fact that it was designed for human listeners should not truly be a disadvantage for speech recognition in practice. With arbitrary sequences humans have as good as no advantage over computers I believe.

My ituition tells me that quite a few would prefer learning the list of words that they might already know partially.

Anyway, my personal ideal spelling feature in a dictation program would show a spelling alphabet of choice in a list with the corresponding letter next to it. This alphabet would likely to be a bit different in each language, since some laguages use letters that are for instance absent from the int'l radiotelephony alphabet eg letters with accents. It's not unlikely that I'd prefer to use the French telephony alphabet, when dictating in French, because that would feel more natural.

--- (Edited on 11/8/2009 5:19 am [GMT-0600] by Robin) ---

Hi Robin,

Sorry for the delay in getting back to you one this, the whole family has been sick at one time or other in the past 3-4 weeks, and the forum posts/emails have been collecting  :)

>can imagine that an application to input text using spelling is not as

>unlikely at all.

You should be able to prototype this in Simon, and see what kind of interest it might generate using the VoxForge acoustic model (or your own...) since the an4 dataset contains spelling recordings.   

And if there was enough interest, then we might look at collecting more of such speech.

Ken

--- (Edited on 11/16/2009 1:58 pm [GMT-0500] by kmaclean) ---