Sayso puts spoken text into the app you are using: press a shortcut, speak, then stop to paste the result. It is a Rust desktop app with a live transcript overlay, downloadable speech models, a terminology dictionary, and editable configuration files—not a cloud dictation subscription.

The 0.7.1 release, published October 11, changes the built-in cleanup prompts to preserve the speaker’s wording, slang, and contractions. Existing customized styles or base prompts need a reset to pick up the change. That is the developer’s intended behavior, not an independent accuracy result.

  • Setup: Install a release package, download a model, and grant microphone access; macOS also needs Accessibility for insertion. Builds cover Apple Silicon Macs on macOS 14+, Windows x64, and Linux x86_64/aarch64. Linux offers .deb, .rpm, AppImage, Flatpak, and tarball packages.
  • Interface: Toggle dictation with Option+Space on Mac, Ctrl+Space on Windows, or Ctrl+Alt+Space on Linux. Settings live in config.toml; cleanup styles are individual TOML files. The app pastes into the focused window and attempts to restore the clipboard.
  • Recognition versus cleanup: Local speech recognition uses platform-specific engines—WhisperKit/FluidAudio on Mac, transcribe-rs on Windows, and sherpa-onnx on Linux. The Raw style skips AI rewriting. Optional styles can use an OpenAI-compatible endpoint or the Claude/Codex command-line tools; a locally launched CLI does not make its model local.
  • Cost and privacy: GPL-3.0 source and official builds are free; the download site accepts a $0 contribution. Optional cloud providers have separate costs and receive audio or transcript data. Local recognition plus Raw keeps speech and text on the machine, according to the docs. History retains text until deletion and audio for 30 days by default; review those controls before dictating sensitive material.
  • Linux limitation: The desktop support matrix matters. Automatic per-app styles do not work on GNOME or KDE Wayland, and per-website styles do not work on Linux. Some keyboard features need input-device access; joining the input group lets every app running as your user read the keyboard. Linux also cannot warn reliably when the focused field is a password field.

Worth evaluating if you want desktop dictation without a mandatory cloud service and prefer file-based configuration. Package availability and documentation are verified here; dictation quality, latency, and application compatibility have not been tested hands-on.