Saydrop

The problem

Dictating is faster than typing, but the common tools ship the recording to a server. For a note that is irrelevant. For a client conversation, a personnel matter or a password spoken by accident, it is not. And the moment the network goes, the dictation goes with it.

How it solves it

Saydrop transcribes locally. Recognition runs through mlx-whisper directly on the Mac’s neural engine. Audio is buffered in memory only and never written to disk. The entire core flow works without a network.

Operating it is a single hotkey. Double-tap starts the recording, another tap ends it, and the finished text lands in whichever application currently has focus.

What it does

Local is the normal case, not a mode. There is no switch to find first. A dictation reaches the cloud only where that was consented to separately and explicitly.

Language clean-up, swappable. After transcription a second model tidies filler words and sentence structure. That can be a local Gemma or Qwen in its own worker process, or the Gemini API.

Failure never costs the dictation. When the clean-up does not answer, Saydrop hands back the raw transcript. An empty field would be the worse answer.

Licences that hold offline. The licence service is a FastAPI application attached to Paddle webhooks. The app verifies the issued licence through an Ed25519 signature with no network. If the licence server is unreachable, Saydrop keeps working.

A whole delivery chain, not just an app. Marketing site and licence service run on the Galaxy Docker Swarm behind Traefik, the DNS record lives in Ansible, the CI project membership in Terraform.

What it does not do

It runs on macOS only, and only on Apple Silicon. Local recognition lives off the neural engine; on Intel Macs, on Windows or on Linux there is none.

It is in daily use by ourselves today and not publicly released. There is no App Store presence and no public download figure we could claim.

The optional clean-up through Gemini does leave the machine. That is why it is a separate consent and not a default.