Google Just Made Voice Search Faster And More Proficient In Noisy Environments

Google prides itself on its voice recognition tech, and it's no wonder: it's good. I take advantage of it on a daily basis for the sake of saving the hassle of typing on a small keyboard for messaging, and while there are occasional hiccups, I am forever grateful that the tech is as good as it is. And now, it looks like it's just gotten better.

Back in 2012, Google shifted its voice recognition tech from the old-school Gaussian Mixture Model method to more advanced Deep Neural Networks. The benefit to the latter is that it was faster, and probably more accurate. But despite that, an issue still existed: the success of using it in a loud environment was hit-or-miss.

With the adoption of some new acoustic models, called Connectionist Temporal Classification and sequence discriminative training techniques, that problem is being tackled.

I first learned of Google's efforts here at NVIDIA's GTC conference this past spring, although I don't recall the models being mentioned by name. The examples that were provided were very impressive, though, so it's great to see that this has become a reality for the regular user so soon.

The tech behind the improved voice recognition is complicated, as you'd expect, and the fact it's even possible is impressive - a testament to the fact that we're dealing with some truly fast hardware nowadays. Ultimately, the models compute larger chunks than normal and as a result, do computations less often. What helps this work in loud conditions is test data that was generated with the help of artificial noise and reverberation. In effect, Google's thrown a lot at this training data, and both CTC and Recurrent Neural Networks (RNN) help deliver accuracy.

Chances are good that your mobile device is already equipped with the software that takes advantage of this updated tech. If you want to be sure, simply go to the updater of your Android or iOS device and make sure everything Google-related is up to date.