Skip to content

Voice identity
for AI that acts.

Know who is speaking, whether the voice is live, generated or replayed, and whether it can be trusted — in the cloud or directly on the device.

Voice agents act.
Verify the human first.

Voice agents now access accounts, make bookings and move money. A clone built from seconds of public audio passes every knowledge check — and the transcript reads the same either way. Vosco scores the caller from the first seconds of the call, and gates the action, not the conversation.

A woman speaking on her phone at an office desk
“I would like to confirm a transaction.”Vosco verifies the human on the line before the agent acts — catching cloned callers and replayed audio in real time.

Trust the caller, not just the transcript.

Your agent understands what is said. Vosco tells it who said it, and whether a real person is on the line — so it can act with confidence.

Stop clones
Synthetic and cloned speech flagged in milliseconds — retrained as new voice models ship.
Catch replays
A recording of the right customer no longer clears the gate.
Know the speaker
Voiceprint verification against enrolled customers, 1:1 or 1:N.
Whole call
Vosco listens for the whole call and returns a fresh verdict every 3 sec — so it keeps confirming it’s the same caller, with no pause in the conversation.
Less friction
Let everyday turns through; step up only when the action or the risk calls for it.
Vosco for voice agents

Machines can hear us.
They need to know who to trust.

Vosco brings voice identity, liveness and anti-spoofing directly onto robots and intelligent devices — so they make trust decisions locally, in real time, with no network.

A person speaking to a humanoid robot at home
“Unlock and get to work.”Vosco confirms it’s the owner, live, and not a cloned or replayed voice. Instantly, on-chip, with no network.

Give every spoken command an identity.

Vosco runs on the robot’s own chip and checks each command — who said it, and whether the voice is live, replayed or cloned — before it reaches the motion or task layer. Offline, in under 10 ms.

Owner-only
Unlock, start tasks or change settings only for enrolled users.
Roles
Operators, supervisors and guests get different command sets on the same robot.
Replay-proof
A recording played from a phone can’t trigger motion or override a limit.
Second factor
For high-risk actions, a paired phone or watch confirms the speaker.
Explore Physical AI

AI is moving from answering
questions to taking actions.

Once AI can act, voice becomes an identity and authorization problem — not only an interface. Vosco provides the trust layer for both: software agents in the cloud, and machines that hear and act in the physical world. One research pipeline, the same models, two places to run them.

Agentic AI

Vosco API

For voice agents that access accounts, make bookings, transact or trigger workflows. Stream audio in and get identity, synthetic-speech and replay signals back in milliseconds — before the agent acts.

  • Synthetic
  • Replay
  • Speaker identity
  • Trust score
Physical AI

Vosco SDK

For robots and intelligent devices that need to know who is speaking and whether they should act. A 7M-parameter model running entirely on the chip — no network, deterministic latency, no audio leaving the machine.

  • On-device
  • Real-time
  • Offline
  • Trusted-device verification

One trust layer.
Multiple signals.

Speaker recognition, trust scoring and voice insights — trained in-house, benchmarked in public, and returned together in one call.

Speakers
  • Voice vectorA biometric profile unique to each voice
  • 1:1 verificationIs this the speaker they claim to be?
  • 1:N identificationWhich enrolled speaker is this?
Trust
  • LivenessLive human vs. recording or replay
  • GenAIFlags synthetic and cloned speech
  • Audio qualityStudio, phone or noisy channel
Insights
  • GenderEstimated speaker gender
  • AgeEstimated age range
  • EmotionNeutral, happy, sad or angry — on request
  • LanguageSpoken language

From audio to a
trust decision.

Vosco listens continuously and returns a fresh verdict every 3 seconds, with inference under 10 ms — so a switch to a different caller, a recording or a clone is caught mid-call, before the action.

  1. Stream or capture audio — over WebSocket, REST, MCP, or straight from the device microphone.
  2. Receive the signals — speaker, liveness, replay, synthetic and a single trust score.
  3. Apply your policy — allow, challenge, deny or escalate, per action and per risk level.
POST /api/v1/audio/analyze
{
  "summary": {
    "liveness": ["live"],
    "synthesis": 0.97,
    "replay": 0.96,
    "snr_db": 21.5
  },
  "segments": [
    { "start_s": 0, "end_s": 5,
      "labels": ["speech", "live"],
      "age": 34, "gender": "male" }
  ]
}

From the blog

All articles

Every AI that listens should run a trust check before it acts.

Vosco is that check — for voice agents in the cloud and for robots, vehicles and devices on the chip.

  • 3 secNew verdict, all call long
  • < 10 msInference latency
  • 7MParameters · Vosco Nano
  • 100%Offline on-device
  • 1,000 min/moFree every month