Few studies carried out in the social sciences are reported in mainstream media so that it is always interesting to see ones that buck this trend.
The latest example was a small scale study that looked at ‘how well a spectrogram can pick out voice features which would identify the speaker as being alcohol intoxicated’ (1). I read a piece about this in one newspaper (2) and a quick online search revealed it had been reported in many more both in UK and internationally. So what was the study about?
Researchers had shown that spectrograms (basically a spectogtram is visual representation of a sound signal as it varies with time) can be used to differentiate when participants were intoxicated and when they were sober. The study involved 18 volunteers who were asked to record tongue twisters [3] every hour for seven hours, first when sober and then when having drunk a ‘high dose of alcohol’. These recordings were analysed through support vector machine models [4] and, it was argued, differentiated between sober and ‘intoxicated’ speech with a high degree of accuracy.
On reading about the research my first thought was that the findings were blindingly obvious: of course our speech changes when are drunk and this should be easy enough to pick up by ear let alone using voice recognition. But as far as the authors were aware these differences had not been modelled rigorously or at least not modelled with such a high success rate. They were showing what was now possible.
My second thought was how on earth were they given ethical approval for the study? In fact that paper does go into some detail about the screening of applicants and the further steps taken so as not to put volunteers at risk. But, to be honest, it is not the kind of study that I am comfortable with.
So why the fuss?
I think it may have been the tongue twisters – as even if few could understand support vector machine models everyone can understand a tongue twister. Second, it was easy to big the research up – some of the articles wrote of an app that was now already available to identify if you were intoxicated or not and mentioned real world applications such as a car that will not let you start unless your speech is first identified as sober. In fact, the paper does not mention any such app and the authors explain that the research is exploratory (proof of concept) and comes with limitations and caveats. The third reason for its popularity is the heavy emphasis being given to AI application and voice recognition now and this research tapped into a sense of just how amazing computers are.
To sum up, researchers knew what they were contributing to the field and the reporting was overblown but in most respects accurate. However, the prominence given to this one study was not justified,but this was scarcely the researchers fault.
Notes
[1] Suffoletto, B., Anwar, A., Glaister, S., & Sejdic, E. (2023). Detection of Alcohol Intoxication Using Voice Features: A Controlled Laboratory Study. Journal of Studies on Alcohol and Drugs, 84(6), 808-813. [Online]https://doi.org/10.15288/jsad.22-00375
[2] Davis, N. (2003) ‘Tongue-twisters could be used to gauge alcohol-intoxication levels, study finds, Guardian, Nov 2023. [Online] https://www.theguardian.com/society/2023/nov/09/tongue-twisters-used-gauge-alcohol-intoxication-levels-study.
This is one example out of very many easily accessed via a web search.
[3] Tongue-twisters are phrases that are difficult to articulate – more so if you have been drinking, for example ‘She sells seashells by the seashore’ or ‘Susie works in a shoeshine shop. Where she shines she sits, and where she sits she shines’.
[4] In brief, support vector machines are ways of analysing data for classification purposes
Michael Hammond 