UNN scientists teach neural network to recognize alcohol intoxication using voice acoustic spectrum

Scientists from Lobachevsky University and Tomsk State University of Control Systems and Radioelectronics have proposed to distinguish a person in a state of alcohol intoxication from a sober one by the sound spectrum of the voice using a specially trained neural network and machine learning methods.

This kind of technology will be in demand in business for monitoring and assessing the condition of employees whose professional activities involve dynamic speech. The effectiveness of the approach does not depend on the language spoken and the level of background noise.

"Methods for determining alcohol intoxication based on voice pitch or amplitude can become a tool for speech-based monitoring of call centre employees. If someone has communication problems for these reasons, the system will determine this automatically using records of conversations with clients," said the author of the study, Valeria Demareva, Head of the Cyberpsychology Laboratory at the UNN Faculty of Social Sciences.

According to Valeria Demareva, this is one way to improve public safety, especially when it is not possible to test a person's blood or breath for alcohol.

More than 600 individuals participated in the experiment, some of them were in the stage of moderate intoxication. The participants were asked to read tongue twisters, then the audio recordings were decoded using spectral analysis of the voice. The dataset comprised 636 tongue twister recordings collected under two conditions: sober and intoxicated (with 1.5 promille of alcohol in their blood). The data obtained became the basis for training several artificial intelligence models, the best of which learned to distinguish very accurately between drunk and sober speech.

The next step in the work of the researchers is to develop algorithms for determining stress states based on continuous speech.

“This will make it possible to quickly test a job applicant's stress tolerance when hiring, as well as monitor the stress level of employees to prevent critical situations,” Valeria Demareva added.

Lobachevsky University scientists have already received a grant from the Russian Science Foundation for the research project “Development of a tool for early identification of stress”.

The findings of the research on distinguishing drunken and sober speech using artificial intelligence were published in The European Physical Journal - Special Topics.