SafeVoice — a proprietary speech-intelligence engine that flags probable alcohol impairment from just a few seconds of ordinary speech. Non-invasive, zero specialized hardware, needing nothing more than a spoken phrase.
Impaired personnel remain a leading and under-instrumented driver of preventable industrial incidents. Alcohol degrades reaction speed, spatial judgment, and fine motor control long before symptoms become visually apparent. Traditional controls — periodic breathalyzers and visual supervisor checks — are intermittent, invasive, bottlenecks at shift gates, and easy to evade. The result is an operational blind spot: high-risk states remain the hardest to intercept continuously.
Voice is the ideal signal to eliminate this gap. Workers naturally use Push-to-Talk radios, access intercoms, and voice interfaces throughout their shift. Alcohol consistently alters acoustic biomechanics — impairing articulation velocity, phonation timing, prosody, and spectral voice quality — biomarkers that a specialized model extracts from seconds of audio.
SafeVoice deploys in two complementary architectures.
SafeVoice-Ref compares real-time voice against the worker's enrolled sober baseline — ideal for shift-start access screening.
SafeVoice-Solo operates zero-shot with no enrolled baseline, assessing single speech bursts during routine radio communications throughout the shift.
Held-out test split of the German Alcohol Language Corpus, the standard public benchmark, scored by unweighted average recall (UAR). Comparisons with prior art use the per-window metric — a single decision per utterance, no voting — because prior art does not vote.
Rather than classifying audio in isolation, SafeVoice-Ref benchmarks speech against the individual's baseline profile. Eliminating inter-speaker acoustic variance yields an immediate ~5-point gain in Unweighted Average Recall (UAR) — our most impactful architectural milestone.
Both Ref and Solo leverage the same low-level acoustic representations. SafeVoice-Ref maximizes precision via a rapid one-time enrollment, while SafeVoice-Solo prioritizes zero operational friction, screening any speaker immediately without historical data.
Standard practice suggests freezing pre-trained backbones on small corpora. Our empirical sweeps proved the opposite: screening accuracy scaled with fine-tuning depth, jumping ~7 points under full end-to-end backbone adaptation compared to partial layer freezing.
Acoustic biomarkers exist across distinct temporal resolutions. Lower backbone layers retain fine articulatory and vocal tract dynamics, while higher layers model prosodic flow and phrasing rhythm. Multi-level fusion followed by attentive pooling dynamically weights the most diagnostic frames.
We bypassed synthetic white noise in favor of extensive industrial audio curricula — heavy machinery, vehicle rumble, ambient crowd chatter, and structural reverberation. This closed the performance gap in harsh field conditions without degrading clean baseline accuracy.
To prevent transient spikes or throat-clearing from triggering false alarms, the inference pipeline scores multiple overlapping analysis windows per utterance. Applying consensus voting adds 1–2 points of UAR and makes detection exceptionally resilient to real-world acoustic spikes.
Consensus UAR (MV@5), reference-based / reference-free. Published including the conditions where accuracy falls.
SafeVoice does not measure blood alcohol concentration and cannot serve as grounds for removing a worker from duty. It is a first-line screening tool that indicates who warrants attention, for verification by established means. The relevant comparison is not with a breathalyzer but with the alternative in use today: supervisor observation, which performs at roughly the level of untrained listening.
All results come from one corpus — the German Alcohol Language Corpus, recorded from 162 speakers with per-speaker blood- and breath-alcohol ground truth. It is the reference benchmark for this task, but it is a single corpus, and every comparison above is made on its held-out test split.
We publish the conditions where accuracy degrades alongside the headline figures, and we publish what we rejected: auxiliary emotion and blood-alcohol-regression heads (negative effect), partial backbone freezing (about seven points worse), decision-threshold tuning (a symptom fix worth 0.2 points), and speaker-embedding conditioning for the reference-free model — a genuine 0.6-point gain, discarded because it would have required per-user enrollment and destroyed Solo's zero-friction advantage.
Shift-start fitness-for-duty check: a short spoken sign-in against an enrolled baseline
Continuous passive monitoring over routine radio traffic, with no enrollment
A calibrated flag routed to the appropriate supervisor for a human decision
Runs on CPU with under one second of latency per analysis window
Degrades gracefully when demographic context is unavailable
Surfaced through the XENOM platform alongside the site's other operational signals