Voice control your robot can trust.
Vosco runs on the robot’s own chip and tells it who is speaking — and whether the voice is live, replayed or cloned — before it acts on a command. Offline, in under 10 ms, with no audio leaving the device.
Robots need a voice
interface they can trust.
Voice is becoming the primary interface for machines — and the whole loop has to run on the device.
Offline
No cloud round-trip — it works where the network doesn’t.
Instant
A verdict in under 10 ms on the NPU, with fixed response time.
Private
No audio ever leaves the device.
Secure
Liveness and anti-spoofing before the machine acts.
Voice identity, solved —
with a second factor for voice.
Our technology combines multimodal signals, evolving voice profiles, liveness detection and a network of trusted devices. For higher-risk actions, a phone, watch or earbuds independently verify the enrolled speaker and give the machine a signed attestation.
Use Vosco directly on the machine today. Add trusted-device verification when an action needs a stronger level of assurance.
Fits into your existing voice pipeline.
Keep your wake word, speech recognition and language model. Vosco sits alongside them on the same audio and adds the one thing your stack doesn’t have: who is speaking, and whether to trust it.
How it plugs in
- Mic audio goes to bothYour speech recognition and Vosco, in parallel.
- Vosco returns the signalsSpeaker, liveness, replay and synthetic, plus a trust score.
- Your command handler checks the scoreBefore passing the intent to the motion or task layer.
Integrate the
Vosco SDK.
A TinyML SDK for iOS, Android and embedded targets. The 7M-parameter model returns the same signals as the cloud API — speaker, liveness, replay, synthetic — entirely on the chip, with no network. Licensed per device, per year.
- Embed the SDK on iOS, Android or embedded Linux targets — TinyML runtime, no network dependency.
- Enroll operators once. Every command is checked against the voiceprint before the motion controller sees it.
- Set the policy per action — and pair a trusted device for the ones that need a second factor.
On-device footprint
| Model size | 7M params |
| Latency · NPU | < 10 ms |
| Latency · GPU / CPU | 20 / 50 ms |
| Speaker accuracy | 98.4% |
| Connectivity | 100% offline |
| Platforms | iOS · Android · Embedded |
| Licensing | Per device · per year |
Design partner programme: we integrate with your team on your hardware and test in your real environment.
How it works on the chip.
Every step runs locally on the device — under 10 ms on an NPU, with no network and no audio leaving the machine.
- Listen
Speech, captured locally
The robot or device receives speech on its own microphones.
- Verify
A trust score, on-chip
Speaker identity, liveness and replay / synthetic signals combine into one trust score.
- Act
Allow, challenge, deny
Or request stronger verification from a trusted device before a sensitive action.
AllowChallengeDeny
What an evaluation looks like.
From first call to a build running on your robot — tested on your hardware, in your environment.
Scoping call
Your hardware, microphones, use cases — and which commands need protecting.
Evaluation build
Vosco for your target platform, with sample code to wire into your command handler.
On-robot test
Measure accuracy, latency and false rejects in your real environment.
Licensing
Per device, per year, when you move to production.
Voice identity for machines that act.
Tell us about your hardware and use case — we’ll walk you through the Vosco SDK and set up an evaluation build.