CopySpeech
AI text-to-speech for your clipboard. Copy twice. Hear it.
CopySpeech reads your clipboard aloud. Copy the same text twice within 1.5 s, and an AI voice speaks it, with a HUD and word-level captions.
Download · Changelog · Contributing · Issues
Features
- Double-copy trigger. No hotkey needed. The tray and the Play page work too.
- 16 engine backends. Local: Kokoro (default), Piper, Kitten, Qwen3-TTS, Pocket, and any CLI TTS tool. Cloud: OpenAI, ElevenLabs, Cartesia, Fish Audio, xAI, Mistral, Google, Microsoft Azure, Edge (keyless), and any HTTP endpoint.
- Local: Kokoro (default), Piper, Kitten, Qwen3-TTS, Pocket, and any CLI TTS tool.
- Cloud: OpenAI, ElevenLabs, Cartesia, Fish Audio, xAI, Mistral, Google, Microsoft Azure, Edge (keyless), and any HTTP endpoint.
- One-click local installs. Kokoro installs from onboarding. Installs stream progress and support optional CUDA.
- Live captions. Words highlight in sync with the audio on the Play page, the HUD, and the browser companion. Captions need an engine that reports timings.
- Streaming playback. Local engines start speaking after the first audio window.
- Speed 0.25×–4×, pitch, and volume. Voice profiles store each combination per engine.
- Duck other apps (Windows). Other apps get quieter while CopySpeech speaks.
- Text cleanup. Strips markdown, expands URLs, and normalizes Unicode. It handles non-Latin scripts without damage.
- Browser companion. Reads your selection through the app. Chrome, Edge, and Firefox are supported. The page highlights the passage as it plays.
- Control server. A local HTTP API for scripts and tools. See Control server.
- Interface. 9 languages, right-to-left layouts, tray, dark and light themes, and an auto-updater.

Experimental features
Experimental features are alpha. They can change or break between releases.
The Experimental Features (Alpha) switch lives in Settings → Advanced. It gates these features:
| Feature | What it does |
|---|---|
| Global hotkey | Starts a reading from any app |
| Audio effects | Walkie-talkie, 8-bit, and Cyberpunk effects per profile |
| Auto-save profiles | Writes profile edits without a Save step |
| History | Keeps a log of past readings |
| Post-processing | Rewrites text with an LLM before speaking |
| HUD overlay | Floating window with a live waveform and captions |
When the switch is off, these features pause. Their settings and saved readings stay on disk.
Plugin-level experiments:
- DSH partial auto-read (0.1.1). Reads only the first or last paragraphs of each automatic reply.
Not yet verified live, per the plugin notes:
- Hermes follow-scroll playback.
- The explicit-Save settings page in the Hermes plugin.
Install
Get a build from Releases. Replace <version> with the release number. The latest is 0.3.0.
| Platform | Asset |
|---|---|
| Windows | CopySpeech_<version>_x64-setup.exe ( .msi also available) |
| macOS (Intel) | CopySpeech_<version>_x64.dmg (Apple Silicon runs it through Rosetta 2) |
| Linux | .AppImage , .deb , or .rpm |
| Browser companion | CopySpeech-Companion-<version>.zip |
Onboarding asks you to pick an engine and a voice on first launch. Kokoro is the default.
Linux note. Wayland blocks global hotkeys. Use double-copy, the tray, or a compositor keybind that calls the control server.
Browser companion (Chrome).
- Start the desktop app first.
- Extract the companion zip.
- Open chrome://extensions, turn on Developer mode, and choose Load unpacked on the extracted folder.
- Grant per-site access in the extension settings.
On Windows, installed builds register the Chrome host at launch. Edge, Firefox, Linux, and macOS need manual setup.
Plugins
Hermes Desktop plugin (copyspeech-hermes)
Reads each new completed Hermes reply aloud through the running app. It can highlight spoken words and follow playback.
- Create a junction so Hermes loads the repo file in place: New-Item -ItemType Junction -Path "$env:LOCALAPPDATA\hermes\desktop-plugins\copyspeech-hermes" -Target "<repo>\copyspeech-hermes"
Create a junction so Hermes loads the repo file in place:
New-Item -ItemType Junction -Path "$env:LOCALAPPDATA\hermes\desktop-plugins\copyspeech-hermes" -Target "<repo>\copyspeech-hermes"
- Open Settings → Plugins → CopySpeech and turn on Read new replies aloud.
- Turn off Hermes's built-in Read replies aloud. If both run, each reply plays twice.
Changes apply when you press Save. Edits to plugin.js reload in the app.
DSH plugin (copyspeech-dsh)
DSH is the DeepSeek Harness. This plugin adds a sidebar panel, a speaker action on each reply, a selection pill, and a playback bar with pause, ±5 s seek, and stop.
- Ask the DSH plugin manager to run install_bundle with the absolute path to copyspeech-dsh/.
- Open the CopySpeech panel. The status line reads CopySpeech is running.
- Send a message and click the speaker in the reply's action row.
The plugin needs the control server on 127.0.0.1:43117. Full details are in copyspeech-dsh/README.md.
Agent voice
The Pi extension and the Claude Code hook speak final assistant replies. See docs/agent-voice.md.
Control server
A local HTTP API on 127.0.0.1:43117:
| Route | Purpose |
|---|---|
| GET /health | Check that the app runs |
| GET /status | Read the current caption |
| POST /speak | Speak text. Pass wait to block until done |
| POST /pause , POST /resume | Pause or resume one reading |
| POST /seek | Move one reading by a signed step |
| POST /stop | Stop playback |
Details are in docs/profile-engine-settings.md.
Development
bun install
bun run tauri dev
| Command | Purpose |
|---|---|
| bun run dev | Frontend only, on port 5173 |
| bun run check | Svelte and TypeScript checks, then lint |
| bun run test | Frontend tests (Vitest) |
| bun run format | Format the repo |
| bun run tauri build | Production build |
| cd src-tauri && cargo test | Rust tests |
| bun run bump | Patch version bump ( bump:minor and bump:major also exist) |
| Layer | Stack |
|---|---|
| Backend | Rust, Tauri 2 |
| Frontend | Svelte 5, SvelteKit, TypeScript, Tailwind CSS 4 |
| Audio | rodio, Web Audio ( OfflineAudioContext ) |
| Tooling | Bun, Vite+ ( vp ), shadcn-svelte |
| i18n | svelte-i18n, Crowdin |
Read Understanding the Codebase before you change code. docs/engines.md explains how local engines connect.
Roadmap
The next release targets low time-to-first-audio and a small VRAM and RAM footprint. The goal is local engines that run well on low-VRAM GPUs and CPU-only machines.
Contributing
Read the Contributing Guide for setup, style, and pull request steps.