Talk to the algorithm – should that be alcorithm? – before getting behind the wheel.
The Back Page is, well, back, but keen to prolong the silly season with another booze story.
Wouldn’t it be a great improvement if instead of using expensive breathanalysers to tell when someone’s drunk, there was an AI for that?
Apparently, it would, and researchers from the Centre for Alcohol Policy Research at La Trobe University are on it. A team led by Abraham Albert Bonela has developed an AI called the Audio-based Deep Learning Algorithm to Identify Alcohol Inebriation (ADLAIA), which can detect pissedness just by listening to a person speak.
We don’t mean listen to what they say. Even a human can reliably discern intoxication when their uncle confides at Christmas lunch that actually Pauline Hanson has a point.
The team trained the AI using the publicly available German Alcohol Language Corpus dataset, which consists of 12,000 audio clips of adults speaking while sober and while drunk, i.e. with blood alcohol level above 0.05%. (We like to think the people were made to say things like Fischers Fritz fischt frische Fische, frische Fische fischt Fischers Fritz and Sie stellte das Tschechische Streichholzschächtelchen auf den Tisch, auf den Tisch stellte sie das Tschechische Streichholzschächtelchen.)
Their resulting model, which requires a 12-second speech snippet to work from, achieved 77% sensitivity and 63% specificity. You wouldn’t give it to the police just yet. But with improvements, the authors write, the program could be used “in emergency rooms, sports stadiums, night clubs, restaurants, and bars, in which an instant identification of inebriation is useful, but breathalysers are often always available as backup devices”.
Because it can run on smartphones, they propose a more annoying use: “Statistics such as how many times the individual was inebriated during the week, month, and other inebriation-related statistics could be included within an application to inform individuals on their drinking history.” I’m good, thanks.
The authors say combining speech patterns with other data such as eye redness might enhance the algorithm’s accuracy, and concede the limitations of their small and exclusively German training dataset.
To collect a larger dataset, we humbly propose they release it as a free pub game app that punters have to say a tongue twister into every time they finish a beer – a session for science.
If you can say “Der dicke Dachdecker deckt Dir dein Dach, drum dank dem dicken Dachdecker, dass der dicke Dachdecker Dir Dein Dach deckte” after three beers, penny@medicalrepublic.com.au salutes you.
