Privacy
Your voice stays yours
How SayIt handles speech, text, and data — in plain language, matching the application's actual behavior.
Recognition runs locally
Speech recognition runs on your device with a locally stored Sherpa-ONNX model. Your audio is processed on your machine, not sent to a speech service.
No account. No API key.
Core dictation works without signing up, without an API key, and without a cloud speech service. The application never asks for one to type by voice.
Enhancement is optional and off by default
Transcript enhancement — via a local Ollama instance or a cloud provider — is opt-in. If you enable a remote provider, transcript text (never audio) is sent to that provider.
Context detection reads metadata only
SayIt detects the active application using process and window metadata. It never takes screenshots, reads your clipboard, logs keystrokes, or scans page and editor contents.
Learning is explicit
Vocabulary rules are created only by you, or when you deliberately choose "Remember" after correcting a transcript. There is no silent learning, and vocabulary never leaves your device.
Safe zones
Applications you mark as protected — password managers, banking — block recording before any audio is captured.
History is off by default
Persistent transcript history is opt-in. The most recent transcript is kept in memory so it survives a failed insertion, and Clear Data removes settings, history, logs, and downloaded models.
Network access, honestly
SayIt touches the network to download speech models and — only if you switch it on — for optional transcript enhancement. The application contains no telemetry or analytics.
No absolute promises
Because model downloads and optional cloud enhancement involve the network, SayIt doesn’t promise a completely private or completely offline setup — it keeps those parts optional, local-first, and off by default, and says so plainly in its own privacy screen.