Lipi (লিপি)

Lipi (Bengali for "script" or "writing") is a fast, lightweight, local-first voice dictation app and AI writing assistant for the desktop.

Press Alt + R in any app, speak, and press it again. Lipi transcribes your voice, optionally polishes it with an LLM, and pastes the result straight into the window you were working in: your editor, browser, terminal, or chat app.

Website · Download · Features · Quick start · Configuration · FAQ · Build from source · Releasing

---

Features

• Dictate anywhere with Alt + R. Lipi records from any app, then brings your original window back and pastes the text for you.
• Offline or cloud transcription. Run Whisper models locally on your CPU or GPU with no internet connection, or use fast cloud APIs such as Groq, Cloudflare Workers AI, or OpenAI.
• Live dictation (optional, local only). Text appears each time you pause while the microphone stays open.
• AI polish. Automatically fix grammar, change tone, summarize, or turn speech into bullet points with a local or remote LLM. Prompts are plain Markdown files you can edit.
• Mini floating widget with Alt + M. Shrink Lipi into a small always-on-top pill that stays out of your way.
• Private, local history. Every note and transcript is stored on your machine in SQLite and can be searched later.

Download

Stable installers, after the first stable release:

Platform	Download
Windows	Setup, MSI
Linux (x86_64)	deb, AppImage, or sudo apt install lipi

On Debian 12 and Ubuntu 22.04 or newer (x86_64):

curl -fsSL https://raw.githubusercontent.com/debjit/lipi/main/install.sh | bash

The script adds the archive key and source, then installs Lipi. Later upgrades are sudo apt update && sudo apt upgrade.

To add the archive by hand:

sudo mkdir -p /etc/apt/keyrings
curl -fsSL https://debjit.github.io/lipi/lipi-archive-keyring.gpg | sudo tee /etc/apt/keyrings/lipi.gpg >/dev/null
sudo tee /etc/apt/sources.list.d/lipi.sources >/dev/null <<'EOF'
Types: deb
URIs: https://debjit.github.io/lipi
Suites: stable
Components: main
Signed-By: /etc/apt/keyrings/lipi.gpg
Architectures: amd64
EOF
sudo apt update
sudo apt install lipi

Later versions are sudo apt update && sudo apt upgrade. The archive lists stable releases only. Dev builds stay on the GitHub Releases page and are not linked as the latest download. The site at lipi.debjit.in reads latest.json on the release-meta branch.

Quick start

1. Install and open Lipi. The setup wizard checks your RAM and CPU and recommends either local or cloud transcription.
2. Pick a speech engine:
   • Local: choose a runner and a model (start with base or small), and Lipi downloads it for you.
   • Cloud: choose a provider and paste your API key.
3. Optional: in Settings > Preferences, turn on Paste into previous application.
4. Switch to any app, press Alt + R, speak, then press Alt + R again to stop.

The transcript (polished by your active AI preset, if auto-transform is on) is pasted into that app.

Note

If you press Alt + R while Lipi's own window is focused, the text stays in Lipi's editor and Lipi does not switch to another window.

Keyboard shortcuts

Shortcut	Action
Alt + R	Start or stop recording from anywhere
Alt + M	Switch between the full window and the mini floating widget

Configuration

All settings live in the Settings page, which has four tabs.

Speech recognition (ASR)

• Engine
  • Remote API: presets for Groq, Cloudflare Workers AI, OpenAI, a local/self-hosted server, or a custom endpoint.
  • Local engine: whisper_cpu, whisper_vulkan (GPU), or faster_whisper. Each runner can be set up with one click; faster_whisper is installed into its own Python virtual environment.
• Models: download tiny, base, small, medium, or turbo_q8 with live progress, and store them in any folder, including an external drive.
• Memory: local models are unloaded after a period of inactivity (2, 5, 10 (default), or 30 minutes, or never). The tab shows current RAM usage and has a button to unload the model now.
• Request timeout: 3 minutes by default (also the minimum), with quick options for 5 and 10 minutes or a custom value in seconds.

Live dictation

Live dictation is off by default and only works with local engines. Cloud and self-hosted APIs ignore the setting and transcribe once when you stop.

• Turning it on downloads Silero VAD (about 2 MB), a small voice-activity detector, into your models folder.
• Speech is split at each pause (about 0.6 seconds of silence) or every 20 seconds, and each piece is transcribed in order. A spinner next to the timer shows when a piece is being processed.
• Cancelling discards the phrase in progress but keeps any text already shown.
• The main cost is running the speech model at every pause. tiny and base usually keep up; medium and turbo_q8 may fall a few seconds behind and keep a CPU core busy.
• AI polish, copying, and pasting still happen once, after you stop recording.

AI post-processing (LLM)

• Providers: add as many endpoints as you like: Cloudflare Workers AI, Groq, OpenAI, or any OpenAI-compatible server such as Ollama, vLLM, or Speaches. Available models are fetched from each provider's /models endpoint.
• Presets: switch between Grammar Fix, Professional Tone, Casual, Concise Summary, Bullet Points, or your own prompts. Presets are Markdown files with YAML frontmatter, stored in the presets folder inside Lipi's data directory, so you can edit them in any text editor.
• Auto-transform: run the active preset automatically as soon as recording stops.

Preferences

Setting	What it does
Language code	ISO-639-1 code such as en, bn, es, or hi. Leave empty to auto-detect.
Auto-copy	Copy the transcript to the clipboard when recording finishes.
Paste into previous application	Bring back the window you were using and paste with Ctrl + V.
Always on top	Keep Lipi above other windows.
Start on system startup	Launch Lipi when you log in.
Alt + R recording behavior	Create a new note for each recording (default), or append to the open note.
Mini widget recording behavior	Save recordings from the mini widget as new notes (default), or append to the open note.

Diagnostic logs

A live log of transcription attempts, errors, endpoints, and models. Copy All Logs produces a formatted report you can attach to a bug report.

Auto-paste on each platform

Platform	Behavior
Windows	Works out of the box.
Linux (X11)	Works out of the box.
Linux (Wayland)	GNOME shows a Remote Desktop permission prompt the first time Lipi pastes. Click Allow, or install ydotool to skip the prompt.

Wayland stops apps from sending keystrokes to other windows unless you allow it, so Lipi sends Ctrl + V through the desktop portal, which asks for permission. If you'd rather paste silently in the background, install ydotool; Lipi uses it automatically when it is available:

sudo apt install ydotool
systemctl --user enable --now ydotoold

FAQ

Where is my data stored? Notes, transcripts, and settings are kept in a local SQLite database:

• Linux: ~/.local/share/com.lipi.app/lipi.db
• Windows: %LOCALAPPDATA%\com.lipi.app\lipi.db

Nothing leaves your machine unless you configure a cloud speech or LLM provider.

Can I use Lipi completely offline? Yes. Use a local speech engine, and either turn off AI polish or point it at a local LLM such as Ollama.

Which local model should I pick? base is a good default on most machines. Use tiny on low-end hardware or with live dictation, and small, medium, or turbo_q8 if you want better accuracy and have RAM and CPU to spare. The setup wizard suggests a model based on your hardware.

Should Alt + R create a new note or append? Your choice. Change it under Settings > Preferences > Alt + R recording behavior.

Building from source

Prerequisites

• Node.js 18 or later

• The Rust toolchain, installed with rustup

• On Debian or Ubuntu, these system packages:

  sudo apt update
  sudo apt install -y libwebkit2gtk-4.1-dev build-essential curl wget file libxdo-dev \
    libssl-dev libayatana-appindicator3-dev librsvg2-dev libasound2-dev \
    python3 python3-venv zenity

Build and run

git clone https://github.com/debjit/lipi.git
cd lipi
npm install

# Run in development mode
npm run tauri dev

# Run tests and type checks
cargo test --manifest-path src-tauri/Cargo.toml
npm run build

# Build release installers
npm run tauri build

Tech stack

• App framework: Tauri v2 with a Rust backend
• Frontend: React 19, TypeScript, Vite
• Audio: cpal for 16 kHz mono capture and hound for WAV encoding
• Local speech recognition: whisper.cpp (whisper-cli) and faster-whisper
• Voice activity detection: Silero VAD v5 via ONNX Runtime
• Storage: SQLite via rusqlite
• HTTP: reqwest with rustls

Project structure

lipi/
├── src/                     # React frontend
│   ├── App.tsx              # Main UI, editor, settings, shortcuts
│   ├── SetupWizard.tsx      # First-run setup wizard
│   ├── App.css              # Styles, including the mini widget
│   └── main.tsx             # React entry point
├── src-tauri/               # Tauri / Rust backend
│   ├── src/
│   │   ├── lib.rs           # Tauri commands, lifecycle events, window state
│   │   ├── main.rs          # Application entry point
│   │   ├── audio.rs         # Audio capture, resampling, live capture
│   │   ├── vad.rs           # Silero VAD and pause-based segmentation
│   │   ├── engine.rs        # Local Whisper runners and RAM supervisor
│   │   ├── transcribe.rs    # Cloud speech-to-text clients
│   │   ├── llm.rs           # LLM chat-completion client
│   │   ├── env_config.rs    # LLM provider configuration and .env sync
│   │   ├── presets.rs       # Markdown preset manager
│   │   ├── models.rs        # Model downloads and runner setup
│   │   ├── paste.rs         # Window restore and auto-paste per platform
│   │   ├── specs.rs         # Hardware detection and recommendations
│   │   └── db.rs            # SQLite storage
│   ├── Cargo.toml
│   └── tauri.conf.json
└── package.json

Releasing

npm run release -- patch bumps the version, commits, tags vX.Y.Z, and pushes. A version with a hyphen is a dev prerelease. A version without one is stable and updates the site. The full flow is in RELEASE.md.

License

Lipi is licensed under the GNU Affero General Public License v3.0.