A voice model spotted alcohol intoxication from one-second clips with 98% accuracy

In a lab study of 18 adults reading tongue twisters before and after drinking, a machine learning model using voice spectrograms identified intoxication above 0.08% breath alcohol with 98% accuracy, 98% sensitivity and 97% specificity.

Illustration of a man speaking into a smartphone at a bar while colourful sound waves flow from his mouth into the phone.
Share this:

Your voice may give away how much you have had to drink. In a small laboratory study, a machine learning model identified alcohol intoxication from short English speech clips with 98% accuracy.

The study was led by Brian Suffoletto of the Department of Emergency Medicine at Stanford University, with Sean Glaister of Stanford and Ayman Anwar and Ervin Sejdic of the University of Toronto. It appeared in the Journal of Studies on Alcohol and Drugs in the November 2023 issue.

Tongue twisters, hour by hour

Eighteen adults aged 21 to 62, 72% of them men, took part. Each read a randomly assigned tongue twister before drinking, then again every hour for up to seven hours after drinking a dose of alcohol based on their body weight.

The recordings were cleaned and cut into one-second windows. The team trained support vector machine models to compare each person’s sober voice spectrogram, a visual map of the frequencies in speech, with their voice at each later time point. Intoxication was defined as a breath alcohol concentration above 0.08%.

Near-perfect detection in a small sample

The models predicted intoxication with 98% accuracy, with a 95% confidence interval of 97.1% to 98.6%. Sensitivity, the share of intoxicated clips correctly flagged, was 0.98, and specificity, the share of sober clips correctly cleared, was 0.97. Positive and negative predictive values were 0.97 and 0.98.

The researchers see a practical use: phones and smart speakers could pick up voice changes linked to drinking and deliver help at the right moment. Until now, they note, data supporting that approach in English had been lacking.

The authors call it a small, controlled study and say larger studies with more varied voice samples are needed to validate the models. Each model also compared a person’s speech with their own sober recording, so the test was of change within one voice rather than of strangers.

Study Details:

  • Title: Detection of Alcohol Intoxication Using Voice Features: A Controlled Laboratory Study
  • Authors: Brian Suffoletto, Ayman Anwar, Sean Glaister, Ervin Sejdic
  • Journal: Journal of Studies on Alcohol and Drugs
  • Publication Date: November 2023
  • DOI: 10.15288/jsad.22-00375