Google Voice Gets Neural Network-Powered Transcription to Slash Errors

By Jamie Condliffe on at

Google Voice has an incredibly useful function that provides you with a transcribed version of your voicemail — but if often gets things wrong. Now, Google is throwing neural networks at the problem to help improve its performance dramatically.

The service now uses Google’s recurrent neural networks, similar to those used by Google’s voice search in Chrome and Now. How does it work? A bit like this:

The neural network chops up the speech it’s hearing into short segments, then identifies vowel sounds. The subsequent layers can work out how the different vowel sounds fit together to form words, words fit together to form phrases, and finally infer meaning from what was just mumbled.

The neural nets have apparently helped Google to slash the number of transcription errors made in Voice by a pretty massive 49 per cent. So there will be fewer piss-takes in your massages from now on. I mean, mistakes in your messages. (Sorry.) [Google]

Image by Salim Fadhley under Creative Commons