Developers

one callcheck_voice_command()

Send the audio and what the agent understood. Get back a decision and the evidence.

Ask before you act.

One call between hearing and acting, as a REST endpoint, an SDK, or a tool your LLM calls on its own.

VerbriskPreview API

Three steps
to a trusted command.

The enterprise API below is in preview: names and fields may change before general availability. The open-source engine is stable today.

Enrol the speaker

Three natural sentences, about two seconds each, become a voiceprint. The audio is not kept.

POST /v1/speakers
files: utterance_1..3.wav
→ { "speaker": "user_0042" }

Check the command

Send the command audio, the wake-word segment if you have it, the transcript and the action's sensitivity.

POST /v1/check
action=critical
→ { "decision": "block", … }

Act on the answer

Obey runs it. Confirm asks the user out of band. Block refuses and keeps the record. That is the whole integration.

obey    → execute
confirm → ask the user
block   → refuse, log

API reference

Preview endpoints. Every response carries the decision, the trust level, each check's score and band, the policy version and the latency.

Endpoint	What it does	Notes
POST/v1/check	Runs every check on one command and applies the policy for its action.	audio, wake_audio, transcript, alt_transcript, action, speaker
POST/v1/speakers	Enrols a speaker from a few utterances.	Stores the embedding only, never the audio
GET/v1/decisions	Lists decisions with filters for agent, outcome and time.	Paged; exportable
GET/v1/decisions/{id}	One decision with its full evidence.	Scores, thresholds, policy version
POST/v1/policies	Publishes a new policy version.	Dry-run with ?dry_run=true
POST/v1/webhooks	Streams blocked and confirm decisions to your systems.	Signed payloads

Agent tool call

Let the model ask for itself.

Register check_voice_command as a tool. Your LLM agent calls it before any action that came from speech, and follows the answer. Works with any model that supports tool use.

The schema is the same one the open-source engine ships
The answer is written for the model: a decision and a one-line reason
The full evidence goes to your audit log, not into the prompt

tool schema

Open source

Start with the engine.

EchoGuard is the open-source signal engine under Verbrisk: the detector, the trust checks, the benchmarks and a local console. Run it on your own audio today.

terminal

git clone https://github.com/swarnali2k17/echoguard
cd echoguard && pip install -e ".[dev]"
echoguard scan clip.wav
uvicorn app.server:app --port 8000   # local console