Inside The First Audio Library of Alcohol-Addled Speech (Which Just Might Help Stop Drunk Driving)

Since 2007, German researchers have been getting subjects drunk and recording them.

Inside The First Audio Library of Alcohol-Addled Speech (Which Just Might Help Stop Drunk Driving)
[Photo: Caiaimage, Martin Barraud, Getty Images]

You get into your car at 3 a.m., sometime in the not-so-distant future, and it won’t turn on. It senses something…off. It asks how you’re doing (okay), what you ate for dinner (good-but-not-great pad thai) and whether the Leafs won or lost that night’s hockey game (the latter, but what else is new). And by this point, your car doesn’t even have to ask, because it already knows you’re drunk. It offers to call you a cab, doing everything short of pouring you into the waiting driver’s back seat.


The only thing it won’t do is let you drive home yourself. And for that, you’ll have the Alcohol Language Corpus (ALC) to thank. Instead of testing drunk drivers after they’re caught in the act, researchers have suggested a more practical, preventative alternative: “The cars themselves listen to the driver, detect that the potential driver is intoxicated, and prevent the car from starting.”

But to train the cars to do so, you’ll first need an audio library of drunk speech.

Between 2007 and 2009, linguistic researchers from the Bavarian Archive for Speech Signals at the Ludwig Maximilian University of Munich and the Institute of Legal Medicine in Munich in Germany convinced 162 men and women to get drunk. Some, only a little drunk. But others, the sort of drunk that causes you to slur your speech, or decide it’s the right time to play “Wonderwall” to party guests on your guitar—or, worse, get behind the wheel of your car when you really, really, shouldn’t. The researchers had each person sit in the passenger seat of a (stationary) car, and recorded their drunken attempts at speech.

The point of all this was to build something that no one else had built before on such a large scale: the first publicly-available audio library of drunk (and sober) speech, the Alcohol Language Corpus. It is, quite literally, a database of difficult conversations available for study. And for the low, low price of just $1,200 plus shipping, you too could own some 30 GB of these recordings on DVD.

“Any person in a high-risk situation, you want to make sure they’re not intoxicated or tired. So certainly drivers are one, but you also have pilots and cops and firefighters–anyone that you would say, ‘Yes, you shouldn’t be drunk and doing what you’re doing,'” said Andrew Rosenberg, an assistant professor of Computer Science at Queens College, and who leads the school’s Speech Lab.


In 2011, Rosenberg co-authored a paper based on ALC data with researchers from Columbia University. Faced with the challenge of separating sober speakers from those who were impaired, “the question was, could we … make a machine come up with algorithms and features that could sort of approximate, reasonably well, what we as people can do?”

So what do drunk people sound like? When people get drunk, they stammer and stutter and correct themselves. Their voice gets higher. Their words sometimes begin to slur. What Rosenberg and his team found was that you could treat intoxication much like you would another accent. You could train a computer to recognize intoxication and all of its quirks, much as you could a person from Manchester or Goa. It wasn’t perfect, by any means, but it was novel, and it worked.

Naturally, car manufacturers are interested in this sort of thing too. Thilo Koslowski, an automotive analyst at the technology researcher firm Gartner, said that he expects vehicles to be come far more self-aware within the next two years–not just of the environment outside the car, but inside the cabin too.

“Companies are beginning to think of what kinds of sensors they can use to detect the kind of state of the driver,” Koslowski said–including audio sensors, but mostly sensors that measure visual, tactile, odour and behavioral input instead.


How often is the driver taking their hands off the wheel? Did they make make a lane change without looking? Do they smell like booze? “It’s a combination of all of these different things where, if you start looking at them and analyzing them, you can kind of determine what state that the driver is in,” Koslowski said.

In 2007, for example, Nissan unveiled a concept car featuring a drunk-driving prevention system that could detect the amount of alcohol in a drivers’ system via odor and perspiration sensors in the transmission shift knob. A camera monitored a drivers’ eyes for signs of drowsiness, while an on-board computer monitored driving behavior for erratic movements or lane drift.

In an email, General Motors spokesperson Dan Flores said the company is also part of “an industry consortium that is conducting research and analysis on how a breath based and/or touch based system could help identify an intoxicated driver,” and includes pretty much every big-name automotive manufacturer around the world.

“However, we are not familiar with any work that involves speech recognition for this purpose,” Flores wrote. A similar email sent to Ford and Chrysler, two of North America’s other largest automakers, was not returned.


The truth is, this technology is a long ways away from leaving R&D. There are certainly privacy issues to consider, as Koslowski points out. “Is it okay for the self-aware vehicle that you own to kind of spy on you?” he mused. “And that’s where I think a lot of the car manufacturers still are a little bit careful that they don’t want to overdo this, because of that expectation that people have in their car to do whatever they want.”

And the technology has to get better first, too. To the best of Rosenberg’s knowledge, no one yet has claimed the ability to reliably detect someone’s level of intoxication based on speech alone—in large part because voice recognition is hard, even when the speaker is sober. In the study conducted by Rosenberg and the Columbia University team, their drunk-detection algorithm was accurate only 73% of the time.

“That means, you want to go and drive your car, 1 in 3 times we’re going to say you’re drunk. That’s bad for a product,” said Rosenberg. Not to mention a difficult conversation between you and your automotive dealer, too.