Sasayaku (囁く)

A cross-platform voice dictation app with AI-enhanced modes. Runs entirely on your machine — speech recognition via whisper.cpp and text processing via Ollama. No cloud services required. "Sasayaku" means "to whisper" in Japanese.

Features

• Fully local — speech-to-text and AI processing run on your hardware, nothing leaves your machine
• Cross-platform — Linux (GNOME/Wayland), macOS, and Windows
• Voice-to-text transcription using whisper.cpp with GPU acceleration (CUDA, Metal)
• AI-enhanced modes — Email, Note, Prompt, Code, and custom modes with configurable system/user prompts
• Works with Ollama — use any local model (default: qwen3:0.6b for speed) or any OpenAI-compatible API
• Floating overlay — unobtrusive recording/processing/result panel with waveform visualization
• Editable results — review and fix transcription before accepting
• App-aware mode switching — automatically picks the right mode based on the active window
• Clipboard integration — Super mode uses clipboard content as context
• Custom vocabulary — map words and phrases for consistent spelling
• Global hotkey — toggle recording from anywhere (Option+Space on macOS, Alt+Space on Windows/Linux)
• Input device selection — choose which microphone to use

Quick Start

1. Install Ollama

Download from ollama.com and pull a small, fast model:

ollama pull qwen3:0.6b

2. Build & Run

See platform-specific instructions below.

3. Download a Whisper Model

On all platforms, you can download whisper models directly from the Settings UI — just pick a model from the dropdown and click Download.

Available models (smaller = faster, larger = more accurate):

Model	Size	Speed	Quality
Tiny	75 MB	Fastest	Basic
Base	142 MB	Fast	Good
Small	466 MB	Moderate	Better
Medium	1.5 GB	Slow	Great
Large v3 Turbo	1.6 GB	Moderate	Excellent
Large v3	3.1 GB	Slowest	Best

4. Configure

On first run, open Settings and configure:

• API Base URL: http://localhost:11434/v1/ (Ollama default)
• API Key: ollama
• Model: qwen3:0.6b
• Whisper Model: select and download from the model picker
• Input Device: choose your microphone

Platform Setup

macOS

Prerequisites: Xcode, Homebrew

# Install xcodegen
brew install xcodegen

# Clone and build
git submodule update --init --recursive
cd macos
./build.sh

Or manually:

# Build whisper.cpp with Metal
cd whisper.cpp && mkdir build && cd build
cmake .. -DGGML_METAL=ON -DCMAKE_BUILD_TYPE=Release
cmake --build . -j$(sysctl -n hw.ncpu)
cd ../../macos

# Generate Xcode project and build
xcodegen generate
xcodebuild -project Sasayaku.xcodeproj -scheme Sasayaku -configuration Release build

The app runs as a menu bar icon. Press Option+Space to toggle recording.

Config location: ~/Library/Application Support/Sasayaku/config.json

Linux (GNOME/Wayland)

Prerequisites: GTK4, libadwaita, PipeWire, Meson, CUDA toolkit (optional)

# Install dependencies (Ubuntu/Fedora)
sudo apt install build-essential meson pkg-config cmake
sudo apt install libgtk-4-dev libadwaita-1-dev libgio-2.0-dev
sudo apt install libpipewire-0.3-dev libspa-0.2-dev
sudo apt install libcurl4-openssl-dev nlohmann-json3-dev
sudo apt install ydotool  # Optional: for auto-pasting

# Build whisper.cpp
git submodule update --init --recursive
cd whisper.cpp && mkdir build && cd build
cmake .. -DGGML_CUDA=ON -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON
cmake --build . -j$(nproc)
cd ../..

# Build Sasayaku
meson setup build
meson compile -C build
./build/sasayaku-daemon

Set up a keyboard shortcut in GNOME Settings pointing to sasayaku-toggle.

Config location: ~/.config/sasayaku/config.json

Windows

Prerequisites: Visual Studio with C++ workload, CMake, curl

# Build whisper.cpp
git submodule update --init --recursive
cd whisper.cpp && mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
cmake --build . --config Release
cd ..\..

# Build Sasayaku
cd windows
mkdir build && cd build
cmake ..
cmake --build . --config Release

Or use the build script: windows\build.bat

Press Alt+Space to toggle recording. The app lives in the system tray.

Config location: %APPDATA%\Sasayaku\config.json

Modes

Sasayaku processes transcriptions through configurable AI modes. Each mode has a system prompt (sets the AI's role) and a user message template (defaults to {transcript}).

Built-in Modes

Mode	Description
Voice to Text	Raw transcription, no AI processing
Email	Formats speech as a professional email
Note	Converts speech to organized bullet-point notes
Prompt	Turns speech into a well-structured AI prompt
Super	Uses clipboard content as context for processing
Code	Generates or explains code from speech

Custom Modes

Create any mode you need in Settings > Modes. Examples:

Business Japanese translator:

• System: You are a translator. Translate the user's spoken English into formal business Japanese (敬語). Output only the Japanese translation.
• User: {transcript}

Git commit message:

• System: You write concise git commit messages. Given a description of changes, output a conventional commit message.
• User: {transcript}

App-Aware Switching

Modes can auto-activate based on the foreground app. Configure auto_apps in the config file or the Modes editor.

Architecture

┌──────────────────────────────────────────────────┐
│                  Shared C++ Core                 │
│  whisper.cpp │ Audio Processor │ OpenAI Client   │
│  Mode Manager │ Recording Coordinator │ Config   │
└──────────────────┬───────────────────────────────┘
                   │
    ┌──────────────┼──────────────┐
    │              │              │
┌───┴───┐    ┌────┴────┐   ┌────┴─────┐
│ Linux │    │  macOS  │   │ Windows  │
│ GTK4  │    │ SwiftUI │   │  Win32   │
│PipeWire│   │AVAudio  │   │ WASAPI   │
│ D-Bus │    │ Carbon  │   │ GDI+     │
└───────┘    └─────────┘   └──────────┘

The core transcription engine, AI client, mode manager, and config system are shared C++17 code. Each platform provides:

• Audio capture — PipeWire (Linux), AVAudioEngine (macOS), WASAPI (Windows)
• Clipboard — GTK (Linux), NSPasteboard (macOS), Win32 (Windows)
• Auto-paste — ydotool (Linux), CGEvent (macOS), SendInput (Windows)
• Window tracking — GNOME Shell (Linux), NSWorkspace (macOS), Win32 (Windows)
• UI — GTK4 (Linux), SwiftUI + floating NSPanel (macOS), GDI+ overlay (Windows)

Configuration

Example config

{
  "api": {
    "base_url": "http://localhost:11434/v1/",
    "api_key": "ollama",
    "model": "qwen3:0.6b",
    "temperature": 0.7,
    "max_tokens": 2048
  },
  "whisper": {
    "model_path": "/path/to/ggml-small.bin",
    "use_gpu": true,
    "language": "en"
  },
  "modes": {
    "email": {
      "name": "Email Mode",
      "use_ai": true,
      "prompt": "You are a professional email writer...",
      "user_template": "{transcript}",
      "auto_apps": ["thunderbird", "outlook.exe", "com.apple.mail"]
    }
  },
  "vocabulary": {
    "API": "API",
    "Kubernetes": "Kubernetes"
  },
  "default_mode": "voice_to_text"
}

Template Variables

• {transcript} — the transcribed speech
• {clipboard} — current clipboard content (for modes with requires_clipboard: true)

License

GPL-3.0 — see LICENSE for details.

Credits

• whisper.cpp by Georgi Gerganov
• Ollama for local LLM inference