Trigger recording
Use push-to-talk on Linux X11 and macOS, or set up a desktop shortcut on Wayland where global hotkeys are handled by the compositor.
Local voice to text for macOS and Linux
Speak once. Clean text anywhere.
A fully local dictation app with a native overlay, on-device transcription, and polished text insertion across your desktop.
Install in one line
Copy and paste this into your terminal.
curl -fsSL https://raw.githubusercontent.com/cesp99/sussurro/master/scripts/install.sh | bashBuilt for the desktop, not the browser
The product is native where it counts.
Sussurro is designed around a simple desktop loop: trigger recording, speak, watch the state change, and keep working in the same app. The overlay, tray, settings window, and text injection are part of the app itself, not a browser shell dressed up as one.
- Native overlay capsule, tray menu, and settings window on Linux and macOS
- System-wide text injection so you can dictate into any focused app
- Open-source Go codebase with native bindings instead of a browser wrapper
Fast enough to disappear
A compact pipeline with one job: turn speech into finished text.
Transcribe locally
Microphone audio is captured on-device and passed to Whisper.cpp, with the overlay reflecting idle, recording, and transcribing states in real time.
Clean and paste
A fine-tuned Qwen 3 cleanup model removes filler words, handles self-corrections, and pastes polished text back into the app you were already using.
Different functions, one local stack
Everything important is configurable, inspectable, and built for actual use.
Native overlay and desktop controls
The pill-shaped capsule stays visible above your workspace and mirrors recording state. You can open Settings or quit from the tray icon or by right-clicking the capsule.
- Idle dots, live waveform bars, and a shimmer transcribing state
- Settings UI for model downloads, language selection, hotkey changes, and mode switching
- Headless `--no-ui` mode when you want terminal-only usage
Local speech pipeline
Sussurro keeps the full speech-to-text path on your own machine. Whisper handles recognition and the cleanup model refines the result before injection.
- Whisper Small for speed or Whisper Large v3 Turbo for higher accuracy
- Qwen 3 cleanup removes disfluencies, speech repairs, and rough punctuation
- Clipboard + paste injection makes the result system-wide instead of app-specific
Configurable for real setups
The docs expose a practical configuration surface so the app can fit different desktops, languages, and model choices without feeling fragile.
- Config file lookup through `-config`, current directory, and `~/.sussurro/config.yaml`
- Language selection supports explicit codes or `auto` for Whisper detection
- Environment variable overrides via the `SUSSURRO_` prefix
Open model stack, deliberate cleanup
Research-backed transcription quality without shipping your voice to the cloud.
The docs describe a modular architecture: audio capture, Whisper.cpp transcription, Qwen 3 cleanup, clipboard handling, and text injection. That means you can understand the path from microphone to final text instead of trusting a black box.
If you want more control, you can switch models, choose the transcription language, override config through environment variables, or run headless for CLI workflows and scripting.
- Dictate into editors, terminals, chat apps, notes, and forms without changing workflow
- Switch Whisper models depending on whether you want lighter runtime or better accuracy
- Run `sussurro-transcribe` for batch audio transcription with the same local models
- Compile from source and inspect the full pipeline if you need an auditable stack
Setup without guesswork
The docs make the first run path clear across macOS, X11, and Wayland.
macOS
Install, launch from a terminal, then grant Accessibility access when prompted so the global hotkey can work.
Linux X11
Install the GTK, WebKit, and tray dependencies once, then use the built-in global hotkey with push-to-talk or toggle mode.
Linux Wayland
Install `wl-clipboard`, then bind your desktop shortcut to the provided trigger script because Wayland does not allow app-managed global hotkeys.
On first launch Sussurro guides you through downloading the required models. Whisper Small is the lighter option, Whisper Large v3 Turbo is the higher-accuracy option, and the Qwen 3 cleanup model is used to refine the final text.
Documentation map
The important docs are worth surfacing, not burying.
FAQ
Questions people usually have before they install.
Does Sussurro send audio to the cloud?
No. The docs describe a fully local pipeline: Whisper.cpp handles speech recognition on-device and the Qwen 3 cleanup model runs locally before text is injected.
What platforms are supported?
The main app supports Linux and macOS. The overlay, settings window, tray, and headless mode are documented for both, with Wayland needing external shortcut setup.
How do hotkeys work on Wayland?
Wayland blocks application-managed global hotkeys by design. Sussurro works there by having you bind your desktop environment shortcut to the included trigger script or socket command.
Can I change models and transcription language later?
Yes. Settings lets you switch Whisper models, download models, and choose the transcription language. The config file also supports language codes, model paths, and environment overrides.
Is there a CLI for audio files?
Yes. `sussurro-transcribe` is the companion CLI for transcribing local audio files with optional LLM cleanup, output files, language overrides, and debug mode.
Do I need to build from source?
No for normal use. The install script downloads the correct release binary for your platform. Building from source is available if you want to compile the native UI or inspect the full stack.
Read the research
The Sussurro-specific research is public, inspectable, and linked here.
Support the developer
Donate directly to Carlo Esposito to help fund continued open-source work.
Sussurro is fully open source and built around a native desktop experience. If it becomes part of your workflow, donations go straight to the developer maintaining it.
Donate on Ko-fiSolana
7jizbu8GD2EGJyxRGxDsSayGD9CV5hkCmUfeYpwvAiNHEthereum
0xA6EAFb432c3bfF4BB1cBaCF19eABbb1da9F56488Bitcoin
bc1qecdkhsh0fpx2zlp9x0smmme9xgwumltsf3m3phStellar
GCHW7CSWW7VA4UZMPSXDHR5CKLZ5DQREUVCJZFXYCCWWJMXTUJUABUCNFully local. Fully open source.