Detect

The raw liveness sensor, as an API. Send audio, get back live, synthetic, or uncertain. Tuned for real telephony, priced by the minute.

Get early access Is this me?

In one sentence

The single building block the other three products are built on, sold on its own to developers who want a telephony-grade voice check inside their own product.

What it is

An APIone call in, one score out

Who buys it

DevelopersVapi, LiveKit, IVR builders

Direction

Eitherinbound or outbound audio

Status

Early accessmetered beta

Who this is for, and the line vs Network

This is the honest boundary, because it is easy to confuse. It comes down to one thing: who the regulator fines.

Detect API · Product 04

Platforms & developers

Vapi, LiveKit, IVR and voice-agent builders. They are not carriers; the actual call still leaves through a carrier underneath them. So they don't face the per-call fine. Their motive is to offer their own users a safety feature. Lower urgency, self-serve, metered.

MOTIVE → add a feature to their product

Network · Product 03

Carriers

Carriers and telcos own the connection to the public phone network, so they are the ones fined if bad calls get through. Their motive is survival, not a nice-to-have. That is why carriers buy the full screening + records service, not just the API.

MOTIVE → avoid per-call fines

Rule of thumb: if the regulator would fine them, they need Network. If they just want a check inside their app, they need Detect. One nuance: if a platform like Vapi ever becomes a carrier itself, it graduates up to Network.

What you send, what you get

One endpoint. Audio in, a verdict out, with an honest uncertain when the model should not guess.

POST /v1/detect● 200 OK · ~1 s
# request: raw telephony audio (G.711 / 8kHz supported)
{ "audio": "<base64>", "codec": "g711_ulaw" }

# response
{
  "verdict": "synthetic",
  "confidence": 0.93,
  "classes": { "live": 0.05, "synthetic": 0.93, "uncertain": 0.02 },
  "model_version": "db-v7-opus",
  "record_seal": "sha256:9f2ac41e…"
}

Even the raw API returns a sealed record hash, so anything you build on it inherits the same tamper-proof evidence trail the bigger products use.

Pricing

Metered by audio analysed, so you pay for what you use. Honest caveat: this is the layer that has to be refreshed every quarter as new voice-AI models ship. It is a sensor, not a finished defense.

Metered

Per minute

of audio analysed

Live / synthetic / uncertain verdict
Tuned for G.711 telephony
Sealed record hash on every call
Quarterly model refresh included

Volume

Custom

for platforms at scale

Committed-use discounts
Dedicated model endpoints
Co-marketing for embedded partners

Strategic note: keep this product, price it, but do not build the company on it. The API is the depreciating layer; every new text-to-speech release erodes any fixed detector. The durable business is Assess, Shield and Network, where the evidence trail and known-voice verification do not depreciate.

For builders

Drop a telephony-grade voice check into your product.

One endpoint, sealed records, quarterly-refreshed models. Early access is open.

Get early access Are you a carrier? → Network