Speech Technology Magazine

 

SpeechTEK 2018: Is the AI Winter Finally Over?

SpeechTEK 2018 is under way, and according to Brian Garr, the "AI Winter" might finally be over, thanks to DNN.
Posted Apr 9, 2018
Page1 of 1
Bookmark and Share

According to Brian Garr, Chief Revenue Officer, GoVivace Speech Technologies, there have been three or four “AI Winters” over the years, depending on how you count. What does he mean, exactly? In his SpeechTEK session, “How Deep Learning Has Changed the ASR Universe,” since the mid-1950s AI has experienced a few false starts. Booms in excitement would be followed by busts—where the interest in AI would be would dry up until something new and more promising came along.

But, Garr says, that cycle might finally be broken—in part because deep learning has changed everything. In the past, ASR suffered from a few big issues. Machines often couldn’t recognize women’s voices, in part because acoustical models were based on the voices of the (usually male) volunteers in the lab. Similarly, background noise posed a huge problem for ASR.

Today, though, much of these issues have been eliminated by the use of Deep Neural Nets (DNN). Now, Garr says, rather than building acoustic models the alphabet is the training data. Now, you can take a spectrogram and match characters to the visuals within it. “This is what changes the world,” says Garr.

This will continue to change speech recognition as we know it. Even now, he says, Alexa knows in advance what intents it will react to. And if you surprise her, she’ll often respond by telling you she doesn’t understand. But “with DNN you don’t need to know intent in advance,” according to Garr. It will just figure it out. With that, AI and speech technologists can breathe a sign of relief that—even as snow continues to fall well into April—the AI winter might finally be a thing of the past.

Page1 of 1