Google Voice Neural Network Update Slashes Transcription Errors Rate By Nearly 50 Percent

Visual voicemail is a wonderful thing...in theory. The idea of being able to read our voicemails in text form is both convenient and awesome, though in practice, it's not usual to stare at a jumble of words trying to figure out if the caller was drunk or if it was simply a poor transcription. Even Google is not shy of admitting that voicemail transcriptions in Google Voice at times "aren't completely intelligible," and at other times "are humorously intelligible."

You may be thinking, 'Great, you've acknowledged the problem, Google, now fix it!' and the good news is you should see continued improvement in your transcribed voicemails. This is something Google has been actively working on, in part by asking users if they would share some of their real-world voicemails so they could analyze and ultimately improve upon the underlying technology. And it's working.

Google Voicemail

"Thanks to those who participated, we are happy to announce an improved voicemail system in Google Voice and Project Fi that delivers more accurate transcriptions," Google stated in a blog post. "Using a (deep breath) long short-term memory deep recurrent neural network (whew!), we cut our transcription errors by 49 percent."

That's a pretty significant improvement, and to benefit from it, there's nothing you need to do other than continue to use Google Voice as usual (assuming you're using Google Voice in the first place). You should see better transcriptions with far fewer errors from here on out, though it's still not likely to be perfect.

"While this is a big improvement, it is just the beginning and with your input, we will continue improving voicemail transcriptions over time," Google added.

Via:  Google
Show comments blog comments powered by Disqus