Skip to main content
Early tutorial (stub). The flow and node configuration are complete enough to build and test end-to-end. Screenshots and a shareable template link will be added as the template is published.
This workflow never flies the drone. The pilot flies manually on the RC; the workflow only listens and raises alerts. We deliberately keep voice out of the flight-control loop — a spoken command should never move an aircraft.
By the end of this tutorial you’ll have a DJI Mini 3 whose pilot can call out what they see — “pallet left by the north gate,” “person near the fence” — and have each observation logged as a structured alert on the drone twin, hands-free, while both hands stay on the sticks. Built entirely in the visual Workflow editor, no code. This is the voice-reporting sibling of the autonomous site sweep (a VLM watches the feed and alerts on its own). Here a human watches and narrates; the workflow just captures it.

​
The idea

Speech is a fast, hands-free way to file an incident while flying. The workflow transcribes what the pilot says, a model decides whether it’s actionable and classifies it, and only real observations become alerts — idle chatter is ignored. No loop, no controller, no drone command — the aircraft is never touched.

​
Prerequisites

  • A DJI Mini 3 twin, paired and streaming (see DJI Mini 3 Site Sweep · Phase 1). The pilot flies it manually on the RC as normal.
  • A microphone twin that streams audio (a headset mic on the pilot is ideal).

​
Step 1: Create the workflow

Create a workflow (DJI Voice Alerts) and add the microphone and DJI Mini 3 twins. Wire nodes left to right; set inputs on the # (fixed) or </> (expression) tabs using {node-name.output}.

​
Step 2: Capture the voice — Audio Track → Audio Assistant

Speak and open Executions — a run fires and Audio Assistant shows is_speaking: true with a captured speech segment.

​
Step 3: Transcribe — Call Model (speech-to-text)

Output: result = the transcript.

​
Step 4: Classify the observation — Call Model (LLM)

Add a second Call Model node (an LLM). Set Prompt to </> and paste the classifier below; the last line inlines the transcript. It decides whether the utterance is a real observation worth alerting on, and structures it.
Output: result = the JSON verdict.

​
Step 5: Read the verdict — JSON Parser

​
Step 6: Alert only on real observations — Conditional

Wire the true port → Send Alert, so idle chatter never raises an alert.

​
Step 7: Raise the alert — Send Alert

Wire Conditional (true) → Send Alert on the DJI Mini 3 twin, mapping the classifier’s fields. The alert lands in the twin’s Alerts panel with the pilot’s observation, tied to the moment and the aircraft.

​
Step 8: Test

You can test this fully without flying — it’s just voice in, alerts out.
  1. Activate the workflow (Simulate or Live; the drone is never commanded either way).
  2. Speak a real observation: “there’s a toolbox left out by the loading dock.”
  3. Check Executions and the Alerts panel:
  1. Speak idle chatter: “okay, battery looks fine.” → should_alert: false, the Conditional blocks it, no alert fires.
Tune the classifier’s should_alert rules for your site so routine flight banter never trips an alert, and genuine hazards always do.

​
Why voice stays out of flight

The pilot’s job is to fly and watch; the workflow’s job is to remember and report. Keeping voice on the reporting side means a misheard word can, at worst, create a spurious alert — never move the aircraft. The RC remains the only thing that flies the drone.

​
Next steps

Autonomous site sweep

Let a VLM watch the feed and alert on its own — no pilot narration.

Send an email on alert

Chain an alert trigger to email or another integration.

Drones

Pairing, telemetry, and the full command set.

Workflow nodes

Every node used here, with inputs and outputs.