AudioBud

AudioBud is a local-first dictation app for Windows. Hold a hotkey, speak, and AudioBud types the transcript into the focused text field. Audio stays on your machine unless you explicitly enable optional LLM post-processing.

AudioBud is a detached fork of Handy by CJ Pais. It keeps Handy's Tauri, Rust, React, and local transcription base while adding AudioBud defaults, a dark frog/swamp interface, a Windows-first release path, and local model choices tuned for this fork.

• Website: https://jamditis.github.io/audiobud/
• Download: latest release
• Changelog: CHANGELOG.md

Current status

The current release is v0.3.0, packaged for Windows x64. The build is not code-signed yet, so Windows SmartScreen warns on first launch. Choose More info -> Run anyway if you trust the build, or build from source.

Windows is the validated target for this milestone. macOS and Linux code is inherited from Handy and may work, but this fork has not validated those builds yet.

Automatic update checks are disabled for now because the inherited updater still points at Handy's release feed. They should return after AudioBud has its own signed release and updater feed.

How it works

1. Hold the default Windows shortcut, Ctrl+Alt+Space, to record. You can switch from push-to-talk to toggle mode in settings.
2. AudioBud records from your selected microphone and trims silence with Silero VAD.
3. The selected local model transcribes the audio.
4. AudioBud inserts the result into the focused app by clipboard paste or direct typing.

What you can configure

• Shortcuts: transcribe, transcribe with post-processing, raw transcript, and cancel bindings.
• Recording mode: push-to-talk or toggle recording.
• Audio: microphone, output device, input meter, audio feedback, volume, and mute-while-recording.
• Models: Parakeet, Whisper, Moonshine, SenseVoice, GigaAM, Canary, Cohere, and custom Whisper GGML .bin files.
• Text output: language selection where supported, translation where supported, trailing spaces, paste method, clipboard handling, and raw lowercased output.
• Vocabulary: custom words plus deterministic word replacements for names, jargon, and common mishears.
• Personalization (opt-in): on-device learning from your own dictation history -- frequently used words offered as suggestions you accept or dismiss and then applied to later dictations, with view, export, and reset controls for everything it has learned.
• Post-processing: optional cleanup through OpenAI, Anthropic, Z.AI, OpenRouter, Groq, Cerebras, AWS Bedrock via Mantle, or a custom OpenAI-compatible endpoint. API keys stay in local settings.
• History: recent transcriptions, recording retention, retry, and saved entries.
• Advanced controls: autostart, tray icon, overlay position, model unload timeout, Whisper acceleration, ONNX acceleration, GPU selection, logging, and debug paths.

Personalization

Opt in to let AudioBud learn from your own dictation history, entirely on your machine. Turn on Learn from my history and it mines your past transcriptions for words you use often and offers them as suggestions; the ones you accept are added to your dictionary and applied to later dictations. It stays off until you enable it, and you can view, export, or reset everything it has learned at any time -- even while the feature is off.

Models

Parakeet V3 is the Windows default in this fork because it was the best small local engine in the milestone A benchmark. See bench/RESULTS.md.

Engine	Best fit	Notes
Parakeet V3	Default Windows dictation	Fast multilingual ONNX model with DirectML support.
Whisper	Broad language coverage	Small, medium, turbo, and large variants through whisper.cpp.
Moonshine	Small English models	Very fast English-focused options.
SenseVoice	Chinese, English, Japanese, Korean, Cantonese	Good option for East Asian language coverage.
GigaAM	Russian	Russian speech recognition.
Canary	Multilingual and translation	180M Flash and 1B v2 options.
Cohere	Accuracy-first multilingual	Larger and slower, but accurate.

Install

Download the Windows installer from the latest release:

• AudioBud_0.3.0_x64-setup.exe
• AudioBud_0.3.0_x64_en-US.msi

On first run, choose a model if one is not already installed and grant microphone permission when Windows asks.

Build from source

Prerequisites: Rust, Bun, and the platform build tools. On Windows, install Visual Studio 2022 with the v143 toolset, the Vulkan SDK, and Ninja. See BUILD.md for platform notes.

bun install
bun run tauri dev
bun run tauri build

For frontend-only work:

bun run dev
bun run build
bun run lint
bun run test

To re-render the README and website screenshots in dark mode:

bun run screenshots

The screenshot script starts Vite, mocks the Tauri command surface with current Windows defaults, captures screenshots/app-general.png and screenshots/models.png, and refreshes the GitHub Pages image assets. It installs Playwright's Chromium browser on first run if needed.

Command-line flags

AudioBud accepts runtime flags for controlling an already-running instance and for changing startup behavior.

audiobud --toggle-transcription   # toggle recording on or off
audiobud --toggle-post-process    # toggle recording with post-processing
audiobud --cancel                 # cancel the current operation
audiobud --start-hidden           # start without showing the main window
audiobud --no-tray                # start without the system tray icon
audiobud --debug                  # enable verbose logging
audiobud --help                   # list all flags

Remote-control flags are sent to the running app through Tauri's single-instance plugin, then the second process exits.

Manual model installation

Use the in-app downloader when possible. If a proxy or firewall blocks it, install model files by hand.

1. Open Settings -> About or debug mode to find the app data directory.
   • Windows: C:\Users\{username}\AppData\Roaming\tech.amditis.audiobud\
   • macOS: ~/Library/Application Support/tech.amditis.audiobud/
   • Linux: ~/.config/tech.amditis.audiobud/
2. Create a models folder inside that directory if needed.
3. Download the model you want:
   • Whisper small: https://blob.handy.computer/ggml-small.bin
   • Whisper turbo: https://blob.handy.computer/ggml-large-v3-turbo.bin
   • Parakeet V3: https://blob.handy.computer/parakeet-v3-int8.tar.gz
4. Place Whisper .bin files directly in models/.
5. Extract Parakeet archives into models/; the extracted folder for Parakeet V3 must be named parakeet-tdt-0.6b-v3-int8.
6. Restart AudioBud. Installed models appear under Settings -> Models.

Custom Whisper GGML .bin files placed in models/ are auto-discovered. The display name comes from the filename.

Debug mode

Open debug mode with Ctrl+Shift+D on Windows and Linux, or Cmd+Shift+D on macOS. It shows app data paths, logs, keyboard implementation settings, recording buffer controls, paste delay, and other diagnostics.

Project layout

• src/ - React settings UI, onboarding, model selector, update checker, translations, and overlay frontend.
• src-tauri/src/ - Rust app setup, managers, Tauri commands, shortcut handling, audio pipeline, transcription pipeline, history, settings, tray, and CLI flags.
• src-tauri/resources/ - default settings, app resources, and tray assets.
• docs/ - static GitHub Pages site.
• screenshots/ - README screenshots generated by bun run screenshots.
• bench/ - benchmark notes and model-selection evidence.

Acknowledgments

AudioBud builds on Handy by CJ Pais and its contributors. It also uses:

• OpenAI Whisper
• whisper.cpp and ggml
• NVIDIA Parakeet
• Silero VAD
• Tauri

License

MIT. See LICENSE. AudioBud is a fork of Handy, and the original copyright notice is retained.