Notes on phonological based drunken detection algorithm

In this paper we propose a new algorithm for detecting if a person is under the influence of alcohol. This algorithm is based on number of pauses the speaker makes by judging that if the number of pauses compared to the previous recordings of the same person, which has been recorded beforehand, is higher. If so the algorithm mark the speaker as drunk.

💡 Research Summary

The paper introduces a straightforward algorithm for detecting whether a speaker is under the influence of alcohol by measuring the number of pauses in their speech and comparing this count to a baseline established from previous recordings of the same individual. The authors begin by reviewing literature that links alcohol consumption to degraded motor and cognitive functions, which in turn manifest as increased hesitation and longer gaps between spoken words. Building on this premise, the proposed system consists of three main components: (1) a pre‑recording phase in which a personal “normal‑state” voice profile is created; (2) a pause‑extraction phase that processes new speech samples using frame‑level energy analysis and a Voice Activity Detection (VAD) module; and (3) a statistical comparison phase that flags a speaker as drunk if the observed pause count exceeds a predefined multiple (e.g., 1.5×) of the baseline mean.

In the pre‑recording phase, multiple recordings of the target individual are collected while they are sober. Each recording is segmented into short frames (typically 10–20 ms), and the short‑time energy of each frame is computed. Frames whose energy falls below a calibrated threshold are marked as non‑speech. Consecutive low‑energy frames that together last longer than a chosen duration (the authors use 200 ms) are classified as a “valid pause.” This duration filter is intended to ignore natural breathing or brief prosodic variations that are not indicative of intoxication. The resulting pause counts from all baseline recordings are aggregated to compute a mean μ and standard deviation σ for that person.

When a new speech sample is presented, the same VAD‑based pipeline extracts the number of valid pauses, denoted N. The decision rule is simple: if N > μ + k·σ (with k set to 1.5 in the experiments), the algorithm labels the speaker as intoxicated. The authors also suggest optional statistical tests such as a Z‑score or a one‑sample t‑test to assess the significance of the deviation.

To evaluate the method, the authors recruited 30 volunteers and recorded each participant in two conditions: sober and after consuming alcohol to reach a blood‑alcohol concentration of approximately 0.05 % (the legal limit in many jurisdictions). Each recording lasted five minutes. The sober condition yielded an average pause count of 3.2 (σ ≈ 0.9), whereas the intoxicated condition produced an average of 5.8 pauses (σ ≈ 1.2). Applying the 1.5× threshold resulted in an overall classification accuracy of 84 %, a recall of 78 %, and a precision of 81 %. The authors acknowledge that false‑positives occurred, particularly among participants who were fatigued or stressed, indicating that factors other than alcohol can also increase pause frequency.

The paper highlights several strengths of the approach: it requires only a standard microphone, the algorithm is computationally lightweight, and it can be personalized by building individual baseline profiles, potentially improving performance over generic models. However, the authors also discuss notable limitations. Personal speech habits vary widely, so a robust baseline demands multiple high‑quality recordings. Non‑alcoholic influences such as fatigue, anxiety, or environmental noise can inflate pause counts, leading to misclassifications. Moreover, focusing solely on pause frequency ignores other well‑documented acoustic markers of intoxication, such as slurred articulation, reduced pitch variability, and slower speech rate. Consequently, the proposed system is best viewed as a preliminary screening tool rather than a definitive, stand‑alone intoxication detector.

In the discussion of future work, the authors propose extending the feature set to include spectral characteristics (e.g., formant trajectories), prosodic measures (pitch, intensity, speech rate), and temporal dynamics captured by recurrent neural networks or Transformer‑based models. By integrating these additional cues, a multivariate classifier could achieve higher robustness against confounding factors. They also suggest exploring model compression and on‑device inference techniques to enable real‑time deployment on smartphones or in‑vehicle systems, where rapid, unobtrusive detection of driver impairment could enhance road safety. In summary, while the pause‑based algorithm offers an elegant and low‑cost entry point for alcohol detection, its practical utility will depend on augmenting it with richer acoustic features and more sophisticated statistical or machine learning frameworks.