“Hi, My name is Melissa and also i’m calling on behalf of terribly a meeting a technology company at the forefront of, and both the city of entertainment. I wanted to follow up with you to make sure you saw love to hear announcement about their new H T M R html 5.”
That gibberish is what Google Voice showed me as the written transcript of a recent voicemail. Given how good Google’s overall voice recognition algorithms have become, Google Voice — and now Google Fi — transcripts always seemed atrociously bad. Now, however, the company promises that it can cut transcription errors by 49 percent.
Google says it is using “long short-term memory deep recurrent neural networks” to cut down on these errors. Using deep neural networks is, of course, the same technique Google uses for its image recognition algorithms and, among other things, spam filters. Google says it used feedback from Voice users to train the new algorithm.
There is probably a boffin somewhere who can explain the “long short-term memory” part of how exactly this works (and feel free to let us know in the comments, right under the guy whose aunt earned $3,584 on her laptop working from home last week — because Facebook’s spam filters apparently don’t use deep learning…), but what matters is that the transcripts will now be more usual.
Having improved Google Voice transcriptions is going to be a godsend for those of us who get a lot of calls — and don’t usually pick them up. Right now, the transcripts are pretty much useless and you still have to listen to the actual message anyway. If Google can at least create usable transcripts for some of them now, that’ll be a big step forward.