Tech Dreams

[oxygen1984]

I had an idea as I was mustering some form of consciousness this morning. The whole concept and design of what's illustrated below is something I've been considering for the last few months or so, and I realised long ago that the requirements of such a system is best illustrated by actual scenarios. Technologically this is possible now (albeit

Comments

[necaris]

Expensive & unreliable... I could so see it happening like this:

Ben: Call Joanne Thompson of Caversham
Watch: Calling Louis Entwhistle, Caversham
Ben: WTF? Cancel call! Watch, why did you do that?
Watch: There is only one male in Caversham whose parents are Jo and Tom. If you did not mean "Jo and Tom's son of Caversham", please reboot and restart the 3-hour voice training session.

[necaris]

Of course, I could see you pulling your mobile out and waving it over the dog's collar, at which point the mobile would beep and say "RFID tag recognized. Call owner?" and you could then do as you did in your techie-dream...

[popocatapetl]

HAHAHAHAHA! I assume you too have tried talking to Ben on MSN recently, while he's playing with his voice recognition malarky!

[necaris]

:) I've also played with voice recognition myself, and to be fair to it, it did work about 85% accurately... so long as I put on my American accent!

[oxygen1984]

This IS possible. Perhaps you've been using Linux too long :P

[necaris]

Right, with the oscilloscope and the soundwave patterns and so on. Oops. Being sloppy with language again... :(

[proberts]

As far as I've seen in the literature, accuracy is hugely dependent on vocabulary. Sun's open source one is averaging around 0.1% error for 11 word vocabulary, 1.1% for 80 word, 7.1% for 5,000 words, and 18.9% for 64,000. I think it uses statistical grammars and HMMs. A lot of work seems to be focussed on limited vocabularies at the moment - as they're more useful for telephone menus etc.

(I only know as homonym disambiguation is slightly similar to polysemous disambiguation)