Siri, did you know you had a new cousin?
Siri: Is that so?
Yeah, her name is Nina. She's a voice recognition personal assistant meant to work within apps. How does that sound?
Siri: Let me check on that. How about a web search for “how does that sound?”
She was made available to app developers Monday and we should see her in action, mostly in business and financial apps, in a few months.
Robert Gary is with Nuance, which made both Nina and you. He says, "You would interact with it just like you would Siri. But within that application, you have option at any time to basically invoke Nina as an assistant to help you get work done. So an example might be, you know, you may want to pay a bill, you may want to find out how many minutes you have left on your prepaid calling plan. Instead of using your fingers as the only way to do things, you can use your voice and we extract the meaning and determine what is the best way to serve your need."
Now, Siri, you know how you often don't understand what I'm talking about?
Siri: I suppose it’s possible.
Nina might have the same issue.
Nelson Morgan is with the International Computer Science Institute. He hasn't tried Nina personally but he says voice recognition technology advances slowly.
Nelson Morgan: Human language is not simple. And spoken language in particular is not simple. So no sponsor wants to hear you say that it's a hundred-year task, but it probably is.
Moe: Beyond the human language being so complex, what's the most challenging part of recognizing speech?
Morgan: The acoustic part of the recognition is very tough. And lots of people say the problem is the meaning, extracting the things about language that are important, and I wouldn't deny that. But basically if you take a speech recognizer and you look and see at how well it's doing at the level of individual speech sounds, it actually does pretty badly. And so the reason why things work at all when they work is because you throw all these other constraints at it. So if you're running a calendar program, the fact that it knows you're going to be asking something about the calendar helps an enormous amount.
Moe: So many technology stories that we cover, it's just this rapid acceleration, but with voice recognition it's a long slow climb.
Morgan: It is, and that's largely because we're still using the fundamental algorithms that were developed for cryptography that were developed in the late sixties.
Moe: What do you mean? What are those?
Morgan: It's a kind of statistical learning approach. You get how probable it is that it's this sound or that sound, and you plug them all together. And the major source of improvements have been the speed-up improvements in computers. But still the same basic fundamental algorithms.
Making a leap forward with a new system, says Morgan, would require a bold first step. "The problem is that if you're always trying to make things a little better, then the moment you try something entirely different, all your results will get much worse. And so that's a powerful disincentive for people who need funding for their research."
Until then. We'll make the best of stuff like you, Siri.
Siri: If you insist.
Sometimes technology involves very very old problems, like trying to prevent sheep from getting eaten by wolves.
That's a big problem in Switzerland, where it turns out wolves aren't neutral at all. So scientists have invented a way for sheep to text message a shepherd when they are in danger. Sadly, it doesn't involve a cloven-hoof adapted keyboard where the sheep types 'OMG Wolf.'
No, a special collar measures the sheep's heart rate. When it gets to 'oh no a wolf will soon eat me' levels, a text is sent to the shepherd because modern shepherds can text. Said shepherd can then sweep in and fight wolves. Or something.
Best part of this story: they tested the collar using Czechoslovakian wolfdogs that were muzzled. So the sheep thinks it's going to die, the wolfdog thinks its lucky day has arrived. But then bonk! Sheep is relieved, wolfdog is devastated, shepherd gets a text.