SomaTraining
A foundation model for continuous physiology. During training, video, audio, metabolic and lab data label the band stream; at inference it runs on the band alone, turning expensive, sparse measurements into daily estimates with confidence attached.
It reads the body, not the sensor.
Raw signal and multimodal labels in; state, forecast and simulation out.
Raw band signal
PPG, acceleration, skin temperature; plus device records such as heart rate, sleep and HRV.
Multimodal labels
Sports video, live and scene audio, metabolic chamber, blood and omics tests, diet and supplement events, genome.
Current state
How is the body now? Daily sleep, recovery and load scores are already produced, and Soma-ACC recognises the current activity every 30 seconds.
Future change
Forecasts with confidence intervals: where things are heading.
Effect of interventions
If exercise or diet changes, what happens? Needs controlled-intervention data.
Soma-ACC: every 30 seconds, what you are doing.
A wrist-accelerometer activity-recognition model, the first member of the Soma family. Trained, and coming soon as open source under Apache-2.0.
Three axes only
No video, no audio: it reads only the wrist accelerometer and outputs body part, activity, intensity and sleep against a shared activity ontology.
About 800k parameters
Small enough to run on a band, a phone or the edge, yet covering the full map of everyday activity.
Trained on public data, released in public
34 public datasets, 960+ participants and 40,000+ hours of signal; every source is credited, and the release is Apache-2.0 on HuggingFace.
Device-agnostic
Trained across 20+ device types; inference needs no device identity, so bands it has never seen work out of the box.
HuggingFace · somaticai/soma-accComing soon
Every layer has a foundation model. No layer has a person.
Genome, methylation, proteome, single cell, wearable signal: each layer has an excellent foundation model. All of them are trained on many people × one layer × one moment. None is trained on one person × every layer × every day. No foundation model has ever seen one person's complete timeline, and that is exactly the model longevity needs.
Continuous signal is the sequence; multimodal data is the label.
The industry treats omics as the input and wearables as auxiliary. We invert it: on one person only continuous signal is dense enough to be the sequence a model reads, and expensive, sparse measurements such as video, metabolism and lab tests label that sequence.
Continuous signal
The backbone the model reads. One instrument, one calibration, with sub-minute morphology kept.
Multimodal reference measurements
They give the sequence meaning. Without labels, state inference cannot be validated and becomes just another black-box score.
Interventions
Exercise, diet, schedule, supplements and medication: exogenous inputs to be explained.
Start with the densest labels.
Sports video
Every shot is a label. Badminton is under way.
Sports and scene audio
Timelines of songs, cues and ambient sound. Concert sessions are already paired.
Functional-medicine tests and metabolic chamber
Energy expenditure, metabolism and lab markers. The chamber protocol is about to start.
Blood and omics tests
Low frequency, high information; mainly for tuning and evaluation.
High-frequency labels are hard because of time alignment.
Video runs at dozens of frames a second, and the band and camera each keep their own clock. Clock offset, drift, Bluetooth latency, the lack of per-sample timestamps and physiological lag all put labels in the wrong place.
Clock model
Estimate each device's offset and drift against a reference clock.
Sync events
Use events visible on both sides, such as claps or shots, to pin the two timelines together.
Signal-level refinement
Fine-tune near sync events using the signals themselves.
Alignment accuracy is graded: the more accurate, the finer the label. Badminton is currently aligned by hand with about 300 ms of error, so labels stay at the event level, rallies and shot windows.
Somatic State Record: what the model says is not the answer.
Typed output
The time-series model emits typed state objects. The language model only translates them into words and never generates a physiological number.
Detection limit is part of the contract
Every record carries a minimum-detectable-difference field. Below that limit, the language model is contractually unable to say "you improved".
Feedback goes into the harness, not the weights
User ratings and corrections enter an attributable, revocable harness layer while model weights stay frozen. Iteration is faster and results stay traceable.
How do you prove the model is getting better?
With a continuous backbone, a twice-a-year test becomes a daily estimate with uncertainty: less noise, no blank days, and every daily estimate is a prediction of the next reference measurement. When the result comes back, the model is either right or wrong. The honest boundary: the reference remains the gold standard and the model will never be more accurate than it. It can be far less noisy.
Starting with one sport.
Badminton
Event-level labels from sports video train the first professional motion-recognition model: courts without cameras can recognise shots and rallies from the band alone. Transferable to tennis, table tennis, running, strength and rehabilitation training.
Running
A 60-second electrode hold before and after the run gives readiness and recovery; during the run, cadence drift and heart-rate decoupling; across a season, a durability curve. A single wrist cannot measure pronation or left-right balance, and we do not report absolute calories.