Helios

Ask an operator at the end of a night shift how they are doing and the answer is almost always "fine". Self-report is the cheapest wellbeing measurement available and the least reliable exactly where it matters most, in the hours when admitting fatigue carries a cost. Speech carries a second channel that is harder to suppress: timing, pitch, articulation and pause structure shift with respiratory, neuromuscular and autonomic state, whatever the words say.

Helios is thymia's non-clinical voice biomarker model. It takes a single recording — thymia's own documentation puts the minimum at ten seconds in one place and fifteen in another, with three minutes the maximum — and returns six scores describing the speaker's state at the moment of recording: mentalStrain, distress, stress, exhaustion, sleepPropensity and lowSelfEsteem. Nothing is fitted per customer or per deployment: a recording is uploaded to a hosted API, a fixed model runs, and scores come back alongside a transcript.

Helios is the wellbeing half of a two-model line. Its sibling Apollo is a clinical instrument, registered with the UK Medicines and Healthcare products Regulatory Agency as a UKCA Class I medical device. Helios is not a regulated medical device, and thymia's documentation is explicit that its scores must not be presented as diagnostic or predictive of any medical condition.

#Key Features

• Six named dimensions rather than one wellbeing number: exhaustion, defined as tiredness resulting from manual or cognitive labour, is scored separately from sleep propensity, the likelihood of falling asleep given the chance, so workload pressure and circadian pressure can be told apart. Mental strain is the composite.
• Population-matched scoring: five of the six biomarkers are returned twice — bucketed into four levels benchmarked against speakers matched on age, birth sex and language or accent, and as continuous values between 0 and 1. Benchmark quality depends directly on the demographic data submitted with the run.
• Built for repeated measurement: runs are tied to a caller-supplied user label, and thymia recommends a per-user baseline over several sessions before acting on any single point-in-time reading.
• Broad capture conditions: 11 languages, six audio formats, any device with a microphone, and deployment on cloud, on-premises or at the edge.
• Batch or live: recordings can be submitted asynchronously through the REST API, or closely related biomarkers under different names — distress, stress, burnout, fatigue and low self-esteem — extracted mid-conversation through Sentinel, thymia's streaming layer.

#Technical Details

thymia's documentation names a single paper as the research behind Helios — its Interspeech 2025 fatigue work, described below — but publishes nothing on the deployed model itself; what is otherwise public is its interface and the research programme underneath it. thymia describes both of its models as reading the same paralinguistic units, which the company calls neural morphemes, and describes Helios as applying the same foundational science as Apollo in an operational framing. On the Apollo side that foundation is set out in a large validation study: an in-house speech encoder derived from TRILLsson5 producing 1,024-dimensional paralinguistic embeddings, with thin calibrated heads on top, trained once on 63,283 recordings from 21,129 speakers and then applied unchanged to new cohorts.

thymia's published fatigue work is the closest evidence for what the Helios readouts measure. A longitudinal study of 1,197 shift workers in six countries, sampled twice daily for two weeks, found paralinguistic features predicted current sleep deprivation and self-reported sleepiness in held-out data, while the more subjective dimensions of physical and mental exhaustion only became detectable under within-speaker modelling. The Interspeech 2025 follow-up thymia points to for Helios recast that per-speaker adaptation as meta-learning over pretrained 1,024-dimensional embeddings, across 10,286 recordings from 1,185 speakers, with a transformer sequence model reaching an AUC of 0.78 for the fatigued state once six prior labelled recordings from a speaker were available. Neither study reports figures for the deployed six-biomarker product, and thymia publishes no evaluation numbers for Helios itself.

#Applications

The deployments thymia names are fatigue and impairment monitoring for drivers, pilots and other operators; wellness tracking for patients between clinical appointments; longitudinal wellbeing measurement across trial participants; and burnout detection for contact centre staff during live calls. The same scores also run over human-to-agent conversations, letting a voice assistant register that a caller sounds exhausted or distressed and adjust or escalate. Because the input is ordinary speech, none of this needs instrumentation beyond the microphone already in the vehicle, headset or phone.

#Impact

Helios is one of the clearer commercial cases for treating voice as a physiological readout rather than a channel for words, and for reporting several separable wellbeing dimensions off one representation instead of a single risk score. Its limitations are equally clear. No code or weights are published and access is by commercial agreement, so the six scores cannot be independently benchmarked and no accuracy figures for them exist in the literature. Scores are relative to a matched population rather than absolute, and single readings are weaker than trends. Used as thymia specifies — as a monitored trend in a non-clinical setting, never as a diagnosis — it is a low-friction measurement where the usual alternative is asking someone how they feel.