Voice notebook

Record one bounded offline transcript and send a save request to the controller.

Documentation revision 1 · Contract reference API 54 · Installed readiness unknown

Preview offline card · Complete HTML interaction body · Machine-readable recipe

What runs where

Explicit owner tap → 10-second microphone clip → installed Android on-device speech recognizer → editable notebook draft → one answer.
No cloud fallback. The card can capture/transcribe without the controller, but saving the notebook requires the controller or another explicitly authorized storage integration.

Controller integration

Controller persists returned transcript only after the owner answers. Do not store credentials or notebooks in example pages.

Only explicit decision-model/weather requests leave the phone; direct or relay transport still requires a running reachable phone.

Boundaries

Not an always-listening notebook. It captures only a bounded clip on a deliberate tap.
On-device recognition availability depends on Android and installed language resources. Empty transcript is not silence evidence.

Lifecycle

start

Configure card data, POST the generated HTML interaction; this does not automatically capture or poll.

update

Owner explicitly refreshes; controller stores approved data and posts a replacement card as needed, cancelling old pending cards.

pause

The Pause button suspends this card’s controls and suppresses results from a pending refresh. There is no background job in this starter; an in-flight capture may finish. Controller integrations must separately pause their jobs.

stop

The Stop button stops this card, suppresses late results, and requests controller shutdown through one response. Confirm/close the controller job and DELETE the interaction separately. Local stop is not proof that an external process stopped.

Failures and costs

No offline recognizer: not_supported before recording. Unsupported language or capture error: show the actual error and let the owner retry.
With return_audio=false, transcription failure fails the request; there is no audio fallback to upload.
No Jev/Clef call. Local microphone and speech consume battery; keep clips bounded.

Complete synthetic contract examples

These fixtures demonstrate contracts. They are not observations from a phone. Replace image/resource IDs with returned authorized IDs. Load request.json as the complete body of POST /v1/interactions; it contains the full offline HTML.

record

{
  "id": "record",
  "request": {
    "method": "POST",
    "path": "/v1/mic/record",
    "body": {
      "duration_seconds": 10,
      "transcribe": true,
      "return_audio": false,
      "language": "en-US",
      "store": []
    }
  },
  "response": {
    "status": 200,
    "body": {
      "recorded_at": "2026-11-04T10:00:00Z",
      "duration_seconds": 10,
      "format": "m4a",
      "sample_rate_hz": 44100,
      "size_bytes": 81200,
      "transcript": "Buy oats on the way home.",
      "on_device": true
    }
  },
  "evidence": "Synthetic contract fixture, not a device observation."
}

Extensions

Add a separate optional controller summarizer after the owner approves the transcript.
Use the separate owner-enabled wake-word system for hands-free collection; that is not silently enabled by this recipe.