How licensing works
- 01
Tell us your hardware. Target chipset, task (transcription, affect, or both), and expected fleet size. A person replies within one business day.
- 02
Free evaluation build. A compiled build for your target device under a time-boxed license — no card, no commitment. Accuracy and latency are measured on your own audio, on your own silicon.
- 03
License per device. If the numbers hold, production is a flat annual per-device rate with model updates included. Custom builds for your footprint and task are scoped separately.
Free evaluation. No card, no commitment.
Why a lab, not a model zoo
The model family
oruk trains its own speech models. On speech-emotion-bench — 64,384 held-out clips, one identical scoring pipeline across 64 systems — the hosted family reaches 77.6% accuracy (trained in-distribution; protocol and caveats on the methodology page). On-device builds are compressed from the same family and validated against your acceptance criteria during evaluation.
The alternative
Free runtimes exist, and for transcript-only prototypes they are a fine start. What they don’t give you: acoustic emotion and speaking-style signals, tuning for your microphone channel and acoustics, a vendor on the hook for accuracy on your hardware, and license terms that survive an acquisition. If your product depends on the speech layer, that gap is the product.
Free evaluation. No card, no commitment.
Exactly what you are buying
The limits below are stated here so you do not discover them after integrating.
- English only
- Models process English speech. No production multilingual support today.
- Specs are measured, not quoted
- We do not publish generic size, memory, power, or latency figures — those depend on your chipset and task, and are established during the evaluation on your hardware. Be suspicious of vendors who quote them unseen.
- Benchmark numbers are the cloud family
- The 77.6% speech-emotion-bench result below was measured on the hosted model family with one open pipeline (methodology, including the in-distribution training caveat). Compressed on-device variants are validated against your acceptance criteria during evaluation.
- What emotion scores are — and are not
- Calibrated acoustic signals about how speech sounds — not intent, truth, medical, employment, or psychological judgments. See responsible use.
- Data and training
- On-device inference means we never see your audio at all. Across the whole product line, oruk never trains on customer content.
- Prefer hosted?
- The same models run behind a synchronous REST API from $0.0045 per audio minute with a $5 trial credit — the fastest way to validate accuracy before committing to a hardware evaluation. See the cloud API.
Free evaluation. No card, no commitment.
Prove it on your silicon
Send your target hardware, task, and expected fleet size. You get a compiled evaluation build and measured numbers on your own audio — then decide.