dybur

Fast, local, private voice dictation for macOS and Windows.

Talk into any text field. Instantly. Without cloud, accounts, or privacy trade-offs. dybur.com

Features

• 100% Local - Speech recognition runs entirely on your device using state-of-the-art ONNX models
• Universal - Inject text into any application via hotkey
• Private - No cloud, no accounts, no telemetry
• Fast - Sub-second transcription latency
• Multilingual - Supports 25 European languages with automatic detection
• Smart - Automatic punctuation and capitalization
• VAD - Voice Activity Detection filters silence for better accuracy

Installation

Download the latest release from GitHub Releases.

Updates

dybur checks for updates automatically when the tray app starts. You can also run a manual check from the tray menu with Check for Updates....

Updates are installed from the public dybur-update.json manifest attached to the latest GitHub release. The manifest points to the stable platform artifacts:

• dybur-windows-x64.exe for the portable Windows app
• dybur-macos-arm64.dmg for the macOS installer

Downloaded artifacts are verified with SHA-256 before installation. On Windows, dybur exits and a helper process replaces the portable EXE, then relaunches the app. On macOS, the helper mounts the DMG, replaces the installed .app bundle, clears the quarantine attribute on the replacement bundle, detaches the DMG, and relaunches dybur.

Set DYBUR_DISABLE_AUTO_UPDATE=1 before launching dybur to skip automatic startup checks. Manual checks from the tray menu still run.

For local release testing, set DYBUR_UPDATE_MANIFEST_URL to point dybur at a test manifest instead of the public GitHub latest manifest. This override applies to manual tray checks and to automatic startup checks in release builds. Debug builds still skip automatic startup checks.

Usage

1. Launch the app (or run dybur start from CLI)
2. Focus any text field
3. Press Ctrl+Shift+Space (default hotkey)
4. Speak
5. Press the hotkey again to stop (or release if using push-to-talk mode)
6. Text appears in the active field

Recording Modes

dybur supports two recording modes:

• Toggle (default): Press the hotkey to start recording, press again to stop
• Push-to-Talk: Hold the hotkey to record, release to stop and transcribe

You can switch modes from the tray menu (Recording Mode) or by editing the config file.

Voice Activity Detection (VAD)

VAD automatically filters silence and background noise before transcription, improving accuracy and reducing processing time. It uses the lightweight Silero VAD model (~2MB) running locally via ONNX.

VAD is enabled by default. Toggle it from the tray menu or via CLI:

dybur vad          # Toggle VAD on/off
dybur vad on       # Enable VAD
dybur vad off      # Disable VAD
dybur vad status   # Show VAD settings
dybur vad threshold 0.6   # Set speech sensitivity (0.0-1.0)
dybur vad min-speech 250  # Set minimum speech duration in ms
dybur vad silence 1000    # Set silence split timeout in ms

All settings and controls are available from the tray menu or via CLI:

dybur start       # Start background service
dybur stop        # Stop service
dybur status      # Check service health (alias: s)
dybur settings    # Open config file (alias: config)
dybur doctor      # Run diagnostics (alias: diag)
dybur models      # Manage speech models (alias: m)
dybur devices     # Manage input devices (alias: d)
dybur vad         # Toggle Voice Activity Detection

Configuration

Config file location:

• macOS: ~/Library/Application Support/dybur/config.json
• Windows: %APPDATA%\dybur\config.json

{
  "hotkey": "Ctrl+Shift+Space",
  "autoPunctuation": true,
  "sentenceCase": true,
  "silenceTimeoutMs": 1000,
  "model": "parakeet-tdt-v3-int8",
  "clipboardCleanup": true,
  "inputDevice": null,
  "recordingMode": "toggle",
  "vadEnabled": true,
  "vadThreshold": 0.5,
  "vadMinSpeechMs": 250,
  "gpuMode": "auto",
  "streamingEnabled": true
}

Option	Values	Description
hotkey	Key combo	Global hotkey to trigger recording
autoPunctuation	true/false	Automatically add punctuation
sentenceCase	true/false	Capitalize first letter of sentences
silenceTimeoutMs	Number	Minimum silence duration used to split VAD speech segments (ms)
model	Model name	Speech recognition model to use
clipboardCleanup	true/false	Restore clipboard after text injection
inputDevice	Device name or null	Microphone to use; null uses the system default
recordingMode	"toggle"/"push_to_talk"	Recording behavior mode
vadEnabled	true/false	Enable Voice Activity Detection
vadThreshold	0.0-1.0	VAD sensitivity (higher = stricter)
vadMinSpeechMs	Number	Minimum speech duration to keep (ms)
gpuMode	"auto"/"cpu"	Use GPU acceleration when available or force CPU
streamingEnabled	true/false	Enable live preview for compatible streaming models

Models

dybur supports multiple speech recognition models. You can switch models from the tray menu or via CLI:

dybur models list      # List available models
dybur models set       # Select a model interactively

Model	Size	Languages	Description
parakeet-tdt-v3-int8	~670 MB	25	Default. Multilingual transducer, balanced accuracy
nemotron-streaming-int8	~660 MB	English	Low-latency streaming transducer
whisper-large-v3-turbo-int8	~1.1 GB	99	OpenAI Whisper, broad language support
whisper-large-v3-turbo-fp16	~1.6 GB	99	Whisper FP16, higher accuracy

Models are downloaded automatically on first use.

Manual Model Provisioning

For offline or locked-down machines, pre-provision models from a connected machine:

1. On the connected machine, run dybur models download <model-id>.
2. Copy the downloaded model directory into the target machine's models directory:
   • macOS: ~/Library/Application Support/dybur/models/<model-id>
   • Windows: %APPDATA%\dybur\models\<model-id>
3. If VAD is enabled, also copy silero-vad from the same models directory.
4. On the target machine, run dybur models set <model-id> and dybur doctor.

Keep the copied metadata.json file with each model directory; dybur uses it for status and cleanup.

ASR Evaluation

Use the local scoring harness to compare saved model hypotheses across a repeatable sample set:

pnpm eval:asr:manifest benchmarks/asr/example.json --require-duration --require-tags
pnpm eval:asr:manifest benchmarks/asr/<run>.json --config benchmarks/asr/corpus-policy.example.json
pnpm eval:asr benchmarks/asr/example.json
pnpm eval:asr benchmarks/asr/example.json --format json --output benchmarks/asr/report.json --strict
pnpm eval:asr:gate benchmarks/asr/candidate-report.json --config benchmarks/asr/gates/candidate-promotion.example.json

The harness reports WER, CER, median latency, realtime factor, and per-tag summaries. See docs/asr-evaluation.md for the manifest shape, reusable corpus policy, and recommended sample set.

Experimental model candidates such as CoreML Parakeet, MLX Parakeet, Qwen3-ASR, and Moonshine are tracked separately from production model IDs. Use packages/core/src/model-candidates.ts to inspect them, scripts/asr-candidates/ for benchmark wrappers, and docs/model-candidate-evaluation.md for the benchmark workflow.

Release smoke checks live in docs/release-smoke-checklist.md.

Requirements

• macOS: 10.15+ (Catalina or later)
• Windows: 10/11
• Microphone: Required for dictation
• Disk: ~700 MB - 1.6 GB depending on speech model

macOS Permissions

On first launch, macOS will prompt for the following:

• Administrator Password - Only if you choose Install Command Line Tool... from the tray menu
• Microphone Access - Required for voice recording during dictation
• Accessibility - Required for injecting text into applications (System Settings → Privacy & Security → Accessibility)

Development

Prerequisites

• Node.js >= 18.0.0
• pnpm 8.10.0+
• Rust (for building the tray application)

Project Structure

dybur/
├── apps/
│   └── tray/              # Tauri 2.0 tray application
├── packages/
│   ├── cli/               # Node.js CLI (@dybur/cli)
│   ├── config/            # Configuration management
│   └── core/              # Core business logic
└── scripts/               # Build and utility scripts

Setup

# Install dependencies
pnpm install

# Build all packages
pnpm build

# Run tests
pnpm test

# Lint code
pnpm lint

# Type check
pnpm typecheck

Building the Tray App

cd apps/tray
pnpm tauri build

Privacy

• All speech processing happens locally on your device
• Audio never leaves your computer
• No cloud services, no accounts required
• No telemetry or analytics
• Logs contain no speech content

License

MIT