Toak is a high-speed, professional-grade dictation tool for Linux. It combines state-of-the-art Speech-to-Text (Whisper via Groq or Fireworks AI) with LLM refinement (Llama/DeepSeek/GPT via Groq, Together AI, Cerebras, or Fireworks AI) to provide a seamless, articulate, and highly configurable dictation experience.

Built with .NET 10 and compiled to Native AOT, Toak runs as a lightning-fast standalone binary with zero runtime overhead, using a client-daemon architecture for near-zero latency.

🚀 Key Features

Blazing Fast: Uses Groq's API for sub-second Whisper transcription and LLM refinement.
Native AOT: Compiled to a native Linux binary for instant startup and minimal footprint.
Intelligent Refinement: Automatically fixes grammar, punctuation, and technical terms while preserving your voice.
Multi-Provider LLM: Supports Groq, Together AI, Cerebras, and Fireworks AI for text refinement, including Llama 3 and DeepSeek models.
Multi-Provider Whisper: Easily swap between Groq and Fireworks AI for state-of-the-art STT APIs.
Reasoning Capabilities: Optional reasoning effort settings for complex text processing.
Modular Skills: Actionable "System" commands for translation, terminal execution, professional rewriting, and summarization.
Multiple Backends: Types directly into your active window (wtype, xdotool, or ydotool), copies to clipboard, or pipes to stdout.
High-Quality Audio: Native support for PipeWire (pw-record) and FFmpeg for universal compatibility.
Beautiful CLI: Interactive onboarding and configuration powered by Spectre.Console.

🛠 Prerequisites

.NET 10 SDK (for building from source)
Audio Capture: pipewire / pw-record (recommended) or ffmpeg
Typing Backend: wtype (Wayland), xdotool (X11), or ydotool (Virtual Input)
Clipboard: wl-copy (Wayland) or xclip (X11)
API Keys: API Keys for your chosen providers (Groq, Together AI, Cerebras, or Fireworks).

📦 Installation

Toak includes a self-contained installation script that handles the native compilation and setup:

git clone https://github.com/your-repo/toak.git
cd toak
./install.sh

The script will:

Publish the project as a Native AOT Release binary.
Install the executable to /usr/bin/toak.
Install Zsh completions to /usr/share/zsh/site-functions/.
Install and enable the toak.service systemd user service so the background daemon runs automatically.

To remove Toak from your system, simply run:

./uninstall.sh

🎮 Usage

Toak operates as a persistent daemon that manages state and API calls, while the CLI sends commands to it via Unix sockets.

Core Commands

toak toggle: The primary command. Start recording; run again to stop, transcribe, and type/copy/stdout the result.
toak start: Explicitly start recording.
toak stop: Explicitly stop recording.
toak status: Output the daemon's current state (Recording, Idle, or JSON).
toak daemon: Runs the background service manually (usually managed by systemd).
toak onboard: Launches the interactive configuration wizard for providers, models, and backends.
toak discard: Instantly aborts the current recording without performing any transcription.
toak latency-test: Benchmarks your network and API latency to ensure optimal performance.
toak show: Displays your current configuration in a clean table.
toak config <key> <value>: Update settings like llm, whisper, lang, backend, punctuation, or tech.
toak skill: Manage dynamic JSON skills via list, add, or remove.
toak history: Query transcription history (-n <count>, --grep <pattern>, --export <file>, --shred).
toak stats: Display usage statistics like total recordings, latency, and top words.

Global Flags

-p, --pipe: Stop recording and output the finalized text to stdout instead of typing it.
--copy: Stop recording and copy the result to the system clipboard.
-v, --verbose: Enable detailed debug logging for troubleshooting.

🤖 Skills System

Toak includes a robust, data-driven skills system triggered by saying hotwords (like "System") at the start of your dictation. Skills are defined as simple JSON files entirely configurable without modifying the C# codebase.

Default Skills

Skill	Hotwords	Description	Type
Terminal	"System terminal", "System run"	Translates request into a bash command and executes it via `terminal_action.sh`.	`script`
Translate	"System translate to [language]"	Translates your dictation into the target language.	`type`
Professional	"System professional", "System formalize"	Rewrites your text to be articulate and formal.	`type`
Summary	"System summary", "System concise"	Strips fluff and provides a direct, crisp summary.	`type`

Customizing & Adding Skills

Run toak skill add to build interactive custom capabilities. All skills live in ~/.config/toak/skills/.

⚙️ Configuration

Key settings in ToakConfig.cs (managed via toak onboard or toak config):

LlmProvider: Choice of groq (default), together, cerebras, or fireworks.
LlmModel: The refinement model (e.g., llama-3.3-70b-versatile or accounts/fireworks/models/deepseek-v3p1).
ReasoningEffort: Set to none (default) or low for O1-style reasoning.
WhisperProvider: Choice of groq (default) or fireworks.
WhisperModel: The STT model (default: whisper-large-v3-turbo).
WhisperLanguage: Set spoken language (e.g., en, es, fr).
TypingBackend: Choose between wtype, xdotool, or ydotool.
AudioBackend: Choose between pw-record (PipeWire) or ffmpeg.
MinRecordingDuration: Set the minimum recording duration in ms (default: 500).
ModulePunctuation: Toggle automatic grammar and punctuation fixing.
ModuleTechnicalSanitization: Ensures technical terms are formatted correctly.