Machines can hear us. They need to know who to trust.
Humanoids, service robots, vehicles and smart devices increasingly ship with microphones — and with no dependable way to know who is speaking or whether to act. A spoken “unlock and get to work” carries real-world consequences. So does the same phrase played from someone else’s phone.
Voice identity for machines has to answer the same questions as for AI agents — who is speaking, and is the voice live — but under much tighter constraints.
Four reasons the check belongs on the device
- Offline. A robot cannot wait on a cloud round-trip to take an order, and it often works where connectivity does not.
- Instant. Safety-relevant functions need a fixed, predictable response time. On-chip inference avoids network variability.
- Private. No audio has to leave the device — important in homes, workplaces and regulated environments.
- Secure. Liveness and anti-spoofing run before the machine acts, so a replayed or cloned command is rejected locally.
The constraint: running on the chip you already have
On-device only works if the model is small enough for the processor already in the bill of materials. That is why model efficiency matters as much as raw accuracy. Vosco Nano runs at 7M parameters with 98.4% speaker accuracy, under 10 ms on an NPU, 20 ms on a GPU and 50 ms on a CPU — fully offline. We explain how we compare models on this axis in the efficiency index, explained.
Listen, verify, act
- Listen. The device receives speech on its own microphones; nothing is streamed out.
- Verify. Speaker identity, liveness, replay and synthetic signals combine into one trust score on the chip.
- Act. Allow, challenge or deny — or request stronger verification before a sensitive action.
A second factor for sensitive actions
Most commands should stay frictionless. For the few that carry real risk — unlocking a door, moving money, overriding a safety setting — a trusted personal device such as a phone, watch or earbuds can independently verify the speaker and send the machine a signed attestation, locally and offline. We cover the design in trusted devices: a second factor for voice.
The full voice stack, on the machine
Identity is one layer of a broader on-device voice stack that also includes wake word, diarisation, local language models, speech recognition and synthesis. Vosco builds the layers that decide who is speaking and whether to act, and partners for commodity layers such as ASR and TTS. See the stack on the Physical AI page.
Frequently asked questions
Why can’t robots use a cloud API for voice verification?
Robots often operate offline, need fixed response times for safety functions, and should not stream household or workplace audio to the cloud. On-device models avoid all three problems.
How small does an on-device voice model need to be?
Small enough for the processor already in the device. Vosco Nano runs at 7M parameters, under 10 ms on an NPU and about 50 ms on a CPU.
Does on-device voice identity work without internet?
Yes. Vosco Nano runs speaker identity, liveness, replay and synthetic detection fully offline on the device.
Put voice identity on the machine.
The Vosco SDK runs on the chip — offline, under 10 ms. Tell us about your hardware and we’ll set up an evaluation build.