Google Voice Failures: Lost in Transcription
I've been using Google Voice for a few years now, since the days before Google bought the service and it went by the name Grand Central. One of the service's best features is its ability to transcribe voice messages and send them to your inbox, making it easy to keep up with incoming calls without disrupting what you're doing at any given moment. If those transcripts actually made sense most of the time, the service would be phenomenally useful.
Unfortunately, making heads or tails of Google Voice transcripts often requires a lot of guesswork, since the text I receive frequently bears little resemblance to the original message the caller left. In most cases I can at least see who called and get a vague sense of what they wanted. But at least a third of the transcripts I get are so riddled with errors that even the caller's identity is a mystery.
Here are a few examples of Google's handiwork. See if you can guess what the caller is saying, and then click the audio stream below the transcript to hear the actual message.
This first message is a pretty typical mix of accurate transcription combined with utter nonsense. Sure, I understand that I'm being invited to a barbecue, but I have no idea what "God for something like that" is supposed to mean, much less "Lots of there and like a giant bolt lock only style." Without listening to the voice recording, I'd be hard-pressed to respond appropriately to this message.
Fortunately, Google uses different shades of gray to indicate its confidence in the transcription. As a general rule, the more gray you see, the less sure you should be of the content. Of course, when the text is basically gibberish, that's a pretty good indicator too.
When the caller mumbles at all, the challenge is even greater, as in the case below.
"Wholesale of the alright"? I have to admit that I'm not 100 percent certain what that message actually said, but I'm pretty sure Google's guess was way off the mark.
Interestingly, longer messages tend to do a little better than shorter ones, perhaps because they give the service more opportunities to hit the mark. In the message above, Google gets a fair amount of the text right. But without hitting the keyword brewery and a reasonable approximation of the name of the beer my friend asked about, there's absolutely no context from which to guess at the content of the message.
I suppose I could chalk mistranscriptions like "pays love the buried" and "the vitamin of there" up to my buddy's accent, but this message actually fares much better than many from people with typical California "TV Land" accents.
"Hit the macksey on?" Not sure what that means, but I have been meaning to catch Inception. Even without listening to the message, I'd be willing to wager that this caller--who is a colleague of mine (I use Google Voice as my main business number)--isn't actually inviting me to hold her during the show. Misconstruing that voicemail transcript could lead to a meeting with our HR director.
The message above is a pretty reasonable example of Google Voice at its best. I get what the caller is asking about without really reading the message, despite the fact that many of the actual words are wrong. Similarly, I can see at a glance that the call below is an invitation to go rafting.
Yep. After listening to the message, it's clear that the text is mostly right. But what makes this transcript interesting is that it seems to reveal something about Google's transcription algorithm. After the caller mentions that she's going rafting on Sunday, Google Voice hears a word--"Yosemite"--that sounds vaguely similar to "Sunday," so the service appears to have assumed that it's the same word. This sort of contextual relevance pops up frequently in Google Voice blunders, occasionally creating contextually coherent sentences that have nothing to do with the real message the caller left.
Next: Same message, different transcripts--and how you can improve your results.
Same Message, Different Transcripts
An old friend of mine--let's call him Mr. Johnson--phones me frequently just to shoot the breeze. So his Google Voice messages offer an interesting array of transcription errors.
Here we see more of Google's method of inferring text from the available data. Johnson doesn't actually say "Hello Bob" at any point, but Google knows my name is Robert, so it just drops that in since it has no idea what's really being said.
Most of the time, though, Johnson's messages are basically the same: "Strohmeyer. Johnson just calling to say 'hey.' That's it. Talk to you later. Bye." But Google finds a variety of different--and equally inaccurate--ways of transcribing that message.
Please give up, indeed. Or should we?
Help Improve Your Own Transcripts
Though automated voice transcription is far from perfect (and, as we've seen here, far from comprehensible much of the time), it has come a long way in a short time, and Google is continually working to make it better. You can help in the effort by giving Google your feedback with the 'Transcript useful?' boxes at the bottom of each transcript you read.
When a transcript is pretty good and gives you the information you need, click the checkmark box. If it's way off, click the X. Either way, Google will ask you to donate that voicemail to the service to help improve its transcription algorithm.
Naturally, if a message is highly personal, you may not want to give Google permission to use it. But if it's as mundane as those I've included in this story, there can hardly be any harm in sharing it to help make the service better for yourself and everyone else who uses it.
Like most Google services, Google Voice is free. So it's hard to complain when the transcripts aren't as accurate as we'd hope. If you're willing to trade a few odd (and sometimes amusing) mistakes along the way and you don't mind pitching in by donating an occasional voicemail to the project, you can get some real value out of the service. Just don't take any transcript at face value.
Robert Strohmeyer is executive editor at PCWorld. You can follow him on Twitter at @rstrohmeyer.