AI and Data Processing
Draft — not legal advice — pending review by a lawyer. This page describes how the product works today; the reviewed text will replace it before the service is sold.
Where Freqwise uses AI, what each provider receives, what is kept, and why the decision about whether you were right is never made by an AI model.
On this page
The principle
AI helps the product hear you and speak to you. It does not decide whether you were right. Whether a readback is correct, whether the aircraft may move, and every mark and score are decided by fixed rules in the training engine, which give the same result for the same words every time.
The controller you hear is synthetic: a computer-generated voice, not a person. The text of your call comes from automated speech recognition, which can mishear.
Where AI is used
| Component | Provider | Receives | Does not receive |
|---|---|---|---|
| Speech recognition: your audio to text | AssemblyAI, EU endpoint | Your audio while you hold the transmit key | Your name, email or training record |
| Controller voice: the controller's text to speech | Cartesia | The controller's words, which include your aircraft callsign (for example a registration you chose) | Anything you said or typed |
| Interpretation of unusual phrasing (only where the deployment enables it) | Google Gemini on Vertex AI, EU multi-region endpoint | The text of your call, your callsign, the station names, callsigns on the frequency and which elements are expected | Audio, your name or e-mail, the correct values, your scores or history |
What decides correctness
- The training engine compares the elements of your call with what the controller issued, one required element at a time, using published rules and labelled training conventions.
- If interpretation is enabled, it may only suggest how to read unusual phrasing. Its suggestion is validated, and it can never mark a readback correct, move the aircraft or change a score.
- If the audio is unclear, nothing is counted against you and you are asked to say it again.
What is kept, and what is not
- Your audio is not stored by us. It is held in memory only while the transmission is processed.
- The text of your calls is not stored by us. It is shown during the session and then discarded.
- The result of each check, and the value recognised for each element, are stored so your debrief and progress work.
- Content-free diagnostics (timings, outcome categories and word counts; no audio and no words) help us keep it working; their operational detail is cleared after 30 days (by a daily job switched on at launch).
- How long each provider keeps what it receives, and their commitments not to train models on it, are being confirmed in our agreements with them (no such opt-out is sent to them today) and will be stated on the sub-processor page.
What we do not do
- We do not use your voice or your calls to train AI models, and we do not give them to anyone to do so.
- We do not create voiceprints, identify speakers or infer emotion from your voice.
- We do not sell your data or use it for advertising.
- An AI model never decides your score.
Known limits
- Speech recognition can mishear clear speech, accents and aviation terms. The radio log shows you the text it heard.
- The synthetic voice may mispronounce a word; the radio log always shows the controller's text.
- Freqwise is a training aid; see the training disclaimer.
More
The privacy policy covers all personal data, providers and your rights. Questions: the contact form, topic “Privacy”.