Voice firewall that flags scam calls while they are live

Voice firewall from CallerAPI listens to the caller's audio on a live call and returns a scam score, the scam category and the tactics it heard, while the call is still in progress. Your switch, app or agent decides what happens next: warn, record, transfer or hang up. It runs on Twilio, Telnyx or a WebSocket from your own switch, at $0.0039 per started minute.

Get an API key
POST your webhook · X-Voice-Event: verdict live call 
{
  "type": "verdict",
  "kind": "filter",
  "from": "+15559990000",
  "verdict": {
    "score": 0.93,
    "level": "scam",
    "category": "Tax Collection",
    "tactics": [
      "authority_impersonation",
      "threat",
      "untraceable_payment"
    ]
  }
}

34

scam categories, the same list as the CallerAPI complaint data.

12

tactics it can name, from urgency to remote access and one-time code requests.

$0.0039

per started minute of call audio, paid in API credits.

It listens to what the caller says

A number check stops known numbers. A scammer on a fresh number still has to say the script. This layer hears it.

  1. Audio in

    The caller's track arrives from a Twilio or Telnyx media stream, or as mu-law or PCM frames on a WebSocket from your switch.

  2. Scored as it speaks

    Each finished sentence is transcribed and checked for the pretext, the pressure and the payoff of a scam.

  3. Verdict out

    A score from 0 to 1, a level from clean to scam, the category and the tactics. Webhook events are signed with HMAC-SHA256.

The 12 tactics are urgency, secrecy, authority impersonation, remote access, untraceable payment, personal data harvest, one-time code requests, threats, prize lures, family fear, refund bait and callback pressure. Each verdict lists the ones it heard, so an analyst sees why a call scored high.

Where it pays for itself

Voice AI platforms

A scam caller burns tokens and minutes on your agent. Get the verdict early and end the session.

UCaaS and CPaaS

Sell live scam alerts on business lines as an add-on. You already meter minutes, so the price maps cleanly.

Call protection apps

Put it on a forwarding number. The call rings through to the subscriber, and the app gets the verdict while they talk.

Signaling first, then content

Falcon, our open-source robocall mitigation software, screens the INVITE before the call connects. This layer checks what the caller says after it connects. Known spam numbers cost less to stop at the switch with the spam phone number database. A verdict is a probability, not proof, so your policy decides the action.

Test it on one transcript first

Send a transcript to POST /api/voice/scan and read the verdict. No phone number needed. Then point a live call at the stream.

Get an API key