Thoth

Scribe to the gods. Typist to you.

Press a key. Speak. Text appears.

Download for macOS · Download for Linux

Installation · Features · Documentation · Contributing

Voice input on the desktop is usually cloud-dependent, subscription-bound, or a chore to set up. Thoth runs speech-to-text entirely on your machine with GPU acceleration. Press a key in any app, speak, and the text lands at your cursor. Nothing leaves the machine. No subscription, no cloud, no internet required.

---

Features

Local, private transcription Fastest on Apple Silicon: Parakeet on the Apple Neural Engine (CoreML), the recommended default Or whisper.cpp with GPU acceleration (Metal on macOS; CUDA/ROCm/Vulkan on Linux) A cross-platform Parakeet (sherpa-onnx) engine as well Nothing leaves your machine; no telemetry; works offline; voice-activity detection trims the silence	Press a key, speak, paste One toggle hotkey (default F13): press to start, press again to stop Text is inserted at the cursor in any app; your clipboard is preserved Import existing audio too (MP3, M4A, OGG, FLAC, WAV) Recording indicator near the cursor with subtle audio cues
Smart correction Register a name once and its mishearings snap to it (phonetic + spelling), with per-term safety so everyday words are left alone Australian/British spelling from the VARCON database, false-friend safe Optional spoken-number conversion ("twenty three" to 23) Filler-word, whitespace, and punctuation clean-up	Optional AI enhancement Post-process locally with Ollama or any OpenAI-compatible endpoint Built-in prompts (grammar, tone, conciseness, summarise) plus your own Length-constrained so it tidies without rewriting Opt-in and clipboard-context aware
History and export Searchable history with waveform playback Original and AI-enhanced versions side by side Export to JSON, CSV, or TXT Configurable retention; SQLite under the hood	Automation and MCP Opt-in loopback control API (token-authenticated, never network-exposed) A bundled Model Context Protocol server so an LLM assistant can drive the dictionary, settings, and history, or transcribe a file On by default for local assistants; toggle live, no restart

---

Installation

After installing, Thoth checks for updates automatically and installs them in-app.

macOS

macOS will block the app the first time you open it because it isn't from the App Store. This is normal and only happens once.

1. Open the .dmg and drag Thoth to Applications
2. Right-click (or Control-click) the app and choose Open
3. Click Open in the dialogue that appears

Alternative: remove the block from Terminal

xattr -dr com.apple.quarantine /Applications/Thoth.app

Linux

1. Download the .AppImage (or .deb) from the latest release
2. Make it executable: chmod +x Thoth_*.AppImage
3. Run it: ./Thoth_*.AppImage

For GPU-accelerated transcription, install libvulkan1 and your GPU's Vulkan driver; without them Thoth falls back to CPU. See the Troubleshooting guide for Wayland and permission notes.

Once it's running, the Getting Started guide walks you through downloading a model, granting permissions, and your first dictation.

---

Tech Stack

Layer	Choice	Why
Framework	Tauri 2.0	Native performance, small binaries, cross-platform
Backend	Rust 2024	Memory safety, audio performance
Frontend	Svelte 5	Reactive UI with runes
Audio	cpal	Cross-platform audio capture
Transcription	whisper.cpp	GPU-accelerated; Apple Neural Engine and sherpa-onnx options
Database	SQLite	Local persistence with migrations
AI	Ollama	Local LLM enhancement (or any OpenAI-compatible endpoint)
Control API	axum	Loopback HTTP control surface for automation
MCP	rmcp	Bundled MCP server for LLM assistants

---

Documentation

Guide	What it covers
Getting Started	First-run setup: download a model, grant permissions, your first dictation
Personal Dictionary	Custom vocabulary and smart name correction (the canonical registry)
AI Enhancement Prompts	Writing effective prompts for the optional Ollama post-processing
Automation and MCP	The control API and MCP server for driving Thoth from an LLM assistant
Troubleshooting	Hotkeys, permissions, paste, GPU, and Wayland gotchas
Product docs	Intent, workflows, and design principles

---

Contributing

pnpm install
pnpm tauri dev    # Development build
pnpm tauri build  # Production build

Requirements

• macOS 14.0+ or Linux
• Rust 1.87+ (2024 edition)
• Node.js 20+
• pnpm

Linux GPU acceleration

whisper.cpp supports GPU acceleration on Linux. Choose the backend that matches your hardware:

GPU	Feature Flag	Requirements
NVIDIA	--features cuda	CUDA Toolkit 12.x, NVIDIA drivers
AMD	--features hipblas	ROCm 6.x
Any (vendor-neutral)	--features vulkan	Vulkan drivers (what the release ships)

pnpm tauri build -- --no-default-features --features vulkan   # what the Linux release ships
pnpm tauri build -- --no-default-features --features cuda     # NVIDIA
pnpm tauri build -- --no-default-features --features hipblas  # AMD

Building from source needs the Linux system libraries and the Vulkan toolchain; see docs/development/linux-setup.md for the full dependency list, runtime packages, and display-server notes. If GPU initialisation fails at runtime, Thoth falls back to CPU automatically.

---

Your voice. Your machine. Nothing else.

Named after the Egyptian god of writing and wisdom, the scribe who faithfully records all that is spoken.

Built on whisper.cpp, Tauri, cpal, and sherpa-onnx. Inspired by MacWhisper, VoiceInk, and Spokenly.

_{MIT Licence · Report a bug · Changelog}