Researchers in Canada enlisted a group of study participants to read a tongue twister before consuming alcohol and then each hour for up to seven hours afterward. By analyzing these recordings, they developed a machine learning system capable of determining with 98% accuracy whether the individual reading the text was under the influence of alcohol. The paper was published in Journal of Studies on Alcohol and Drugs.
Alcohol is one of the oldest known psychoactive substances. Its consumption is deeply ingrained in numerous cultural and social practices worldwide. People drink alcohol for fun, to socialize, when they are happy or sad, but also in different cultural or religious occasions. Historically, it has been used to purify drinking water (such as grog provided to 18th-century sailors), as a disinfectant, and even as fuel for cars and rockets, among other purposes.
However, excessive alcohol consumption can lead to a variety of negative health effects. In the short term, it can cause drunkenness, characterized by mood swings, reduced inhibition, impaired judgment and coordination, blurred vision, slurred speech, and difficulties in walking or standing. Prolonged and excessive alcohol use can result in addiction or alcoholism, liver diseases, cardiovascular issues, and an increased risk of certain cancers, such as those of the liver, mouth, and throat.
In their new study, Brian Suffoletto and his colleagues noted that, currently, there are no commercially available tools to unobtrusively and effectively identify alcohol intoxication (i.e., drunkenness). Specialized devices like transdermal alcohol sensors and portable breath alcohol meters can accurately estimate blood alcohol content, but they are often expensive and are not widely available. They can often be too burdensome for widespread practical use.
On the other hand, it is well-known that alcohol alters speech and speech can easily be recorded using widely available everyday devices (e.g. mobile phones, microphones). So, analysis of voice samples could be a very easy and effective way to detect alcohol intoxication, without the need for specialized devices, if such a method existed. These authors set out to develop such a method.
The study involved 20 adults, but analyses were conducted on 18 participants as two did not provide voice samples. The average age of these participants was 29 years, with 72% being male.
On the study day, participants arrived at the laboratory at 8:00 AM. Each was asked to read a tongue twister aloud, which was recorded using a mobile phone. After this initial recording, the researchers administered a quantity of vodka, mixed with lime juice and simple syrup, sufficient to achieve a breath alcohol concentration above 0.20%. The participants had an hour to consume this mixture. Subsequently, every half hour for up to seven hours, the researchers measured the participants’ alcohol levels and recorded them reading a tongue twister.
They used these recordings of participants reading tongue twisters to develop a machine learning model for predicting alcohol intoxication. The final model was 98% accurate in predicting alcohol intoxication. It demonstrated that alcohol intoxication can accurately be predicted by analyzing voice recordings.
“We found in this proof-of-concept lab study that brief English speech samples are useful to classify alcohol-intoxicated states in adults. A much larger participant pool with more varied voice samples collected before and during the ascending and descending curves of alcohol intoxication is urgently needed to move the science of remote alcohol intoxication detection forward,” the study authors concluded.
The study made a valuable contribution to developing ways to measure alcohol intoxication using voice analysis. However, this study used a small sample of English speakers who cooperated with researchers. Results might not be the same if the sample of voices was more diverse or if participants actively tried to conceal their alcohol intoxication, as people are often motivated to do in real-world alcohol intoxication measurements.
The study, “Detection of Alcohol Intoxication Using Voice Features: A Controlled Laboratory Study”, was authored by Brian Suffoletto, Ayman Anwar, Sean Glaister, and Ervin Sejdic.
