Private, local text-to-speech. Say It turns copied text into speech with open models running entirely on your machine — your text and generated audio never leave your computer.
This is a Linux-first port of callebtc/sayit (macOS / Apple silicon). It keeps the original architecture and CLI while replacing Apple-specific layers with Tauri, Svelte, and kokoro-js.
Linux (X11 and Wayland) is built and tested. Windows CI builds an experimental MSI (needs Node and mpv on the machine; playback is not a full Windows port). macOS should compile; help wanted.
- Speak from anywhere. Copy text and press the clipboard hotkey
(Ctrl+Alt+V on X11), or bind
sayit-clipboardas a custom shortcut in your desktop environment — the reliable path on Wayland. - A desktop player. Tray window to speak, pause, seek, change playback speed, and revisit history without leaving your current app.
- Open models. Download supported Kokoro-82M weights in the app or CLI
(
kokoro-q8~90 MB, orkokoro-q4). Speak never downloads on its own. - Efficient model loading. Only one model is kept in memory, and it is unloaded after a configurable idle period (ten minutes by default).
- Local by design. Synthesis works offline after model download. There is no analytics, cloud inference, or passive clipboard monitoring.
- Hear your coding agent work. The bundled Say It agent skill provides live, hands-free spoken progress updates while an agent works.
Two products, one engine. The CLI installer never needs a window. The GUI package includes its own copy of the sidecar and talks to whatever is already on port 7878.
Needs Node ≥ 20, npm, and mpv (aplay is a limited fallback).
Clipboard tools (wl-paste, xclip, or xsel) only if you want the hotkey.
curl -fsSL https://raw.githubusercontent.com/ildella/sayit/master/scripts/install.sh | bash -s -- --systemdOr from a clone: bash scripts/install.sh --systemd. Omit --systemd to
start the daemon once without enabling it. Put ~/.local/bin on your PATH.
Then download a model (once, then the app stays offline) and speak:
sayit models install kokoro-q8 --use
sayit "Hello from Say It"Or copy text and run sayit-clipboard (bind that in your desktop
shortcuts). Speak returns an error until a catalog model is installed
and selected.
The Linux GUI is an AppImage. It embeds the sidecar. Needs Node ≥ 20 and mpv on the machine (same as the CLI). No sudo. Tags also publish an experimental Windows MSI (same Node + mpv requirement; not a full Windows port).
Download it from Releases, then:
chmod +x SayIt-*.AppImage
./SayIt-*.AppImageThe window binary is sayit-desktop; it does not replace the CLI sayit.
You can run both: whoever starts first owns port 7878; the other connects.
Auto-update (Settings → Check for updates) applies to the AppImage and the
experimental Windows MSI. .deb / .rpm stay distro packages.
The install includes a sayit CLI for speech, playback, models, and
automation:
sayit "Read this aloud"
printf 'Read standard input' | sayit
sayit status
sayit pause
sayit resume
sayit volume 0 # silence; 1 = normal, 2 = boost
sayit service status
sayit skill pathRun sayit --help for all commands. The CLI talks to the sidecar on
127.0.0.1:7878. If the daemon is down: sayit service start (or
systemctl --user start sayit after --systemd).
After install:
sayit skill installThat copies SKILL.md to ~/.agents/skills/sayit/ (OpenCode and other
agents that read that directory). Then tell the agent:
Load the Say It skill and use it for live spoken updates.
Claude Code (if you use it instead):
mkdir -p ~/.claude/skills/sayit
cp "$(sayit skill path)" ~/.claude/skills/sayit/SKILL.mdRe-run sayit skill install after upgrading Say It. The sidecar must be
running and a model installed before speech works.
In-app global shortcuts (Ctrl+Alt+V) work on X11. Wayland compositors
block them — bind a custom shortcut in your desktop settings to
sayit-clipboard (installed next to the CLI). There is no cross-compositor
API for another app's selection (not clipboard); on X11 you can point
sayit-clipboard at xclip -o (PRIMARY) instead.
Contributors: live UI. The end-user GUI package is the AppImage.
You need Rust and Tauri's prerequisites as well as Node ≥ 20 and mpv.
git clone https://github.com/ildella/sayit.git && cd sayit
npm run setup # sidecar + CLI into ~/.local (same as install.sh)
npm install # @tauri-apps/cli
npm --prefix app install # SvelteKit UI
npm run dev # live UI (developer loop, not a distribution)If a daemon is already listening on port 7878, Tauri connects to it instead of spawning a second one.
npm run build:appimage # AppImage with sidecar inside (Linux CI)
npm run build:linux # .deb (optional)
npm run build:msi # MSI with sidecar inside (Windows CI; Windows host)
npm run build:ci # compile the shell, skip installers (macOS CI)After pulling updates, re-run npm run setup and restart the service so a
CLI install is not talking to a stale sidecar.
The SvelteKit frontend (in a Tauri v2 tray shell) is separate from a
per-user Node sidecar that owns model downloads, synthesis, playback, and
history. The app, CLI, and sayit-clipboard talk to that service over a
token-protected REST + SSE API bound to 127.0.0.1:7878. There is no
Python; synthesis is Kokoro-82M via kokoro-js / onnxruntime-node (CPU).
┌──────────────┐ REST + SSE, Bearer token ┌──────────────────┐
│ Tauri v2 app │ ◄──────────────────────────► │ sidecar (Node) │
│ SvelteKit UI │ │ kokoro-js engine │
│ sayit CLI │ ◄──────────────────────────► │ mpv playback │
│ sayit-clipboard │ history, models │
└──────────────┘ └──────────────────┘
How the port maps onto the original (XPC → loopback HTTP, MLX → kokoro-js, selection → clipboard), filesystem layout, and invariants are in LINUX.md.
Say It was created by callebtc as a privacy-first macOS app. This project exists thanks to his generosity in releasing it under MIT. For the original Apple-silicon experience (MLX Audio, voice cloning, Voice Studio), use callebtc/sayit.
MIT, like the original. Models are distributed under their own licenses. Only synthesize voices you have the right to use.