Multi-platform · macOS + Windows

VachaVox

Speak into your everyday apps. Keep speech processing on your computer.

One product. Two desktops.

Hold a key. Speak into any app.

Local voice dictation for Mac and Windows. Turn speech into text in the apps you already use, with desktop shortcuts and speech models that run on your computer.

macOS
Windows
Local desktop app

Get for macOS Get for Windows

Windows x64 preview · unsigned installer.

macOS source

VachaVox macOS menu-bar popover showing dictation controls — macOS / Menu bar

The project in plain English

Why it matters

For People who prefer speaking to typing

The problem

Dictation can interrupt the task at hand or require sending audio to a cloud service.

My contribution

Built desktop editions for macOS and Windows: a shortcut captures speech, a local model turns it into text, and platform integrations deliver the result.

The value

Brings voice input into the apps people already use while keeping the core transcription workflow on the device.

Skills demonstrated

On-device AI
Cross-platform integration
Workflow design

Scope & limits. macOS and Windows use separate desktop implementations. The Windows edition is an unsigned preview with private source. Speech recognition is local; Windows text cleanup is optional and cloud cleanup requires explicit configuration. Review important transcripts.

One local workflow, platform-specific engines

Models & local processing

Both editions run speech recognition on your computer. Use the model format supported by your edition; language coverage, hardware needs, and transcription quality depend on the selected model.

macOS

Parakeet & Whisper / Core ML

Local model folders through FluidAudio and WhisperKit.

Parakeet TDT v2 / v3 and CTC 110M Core ML options.
WhisperKit choices from Tiny through Large v3 / Turbo, plus Distil-Whisper.
Install compatible model folders separately, then select and load one in Settings → Models.

Model installation guide ↗

Windows

Whisper & native inference

The 0.1.2 preview bundles verified Whisper Base (English).

Other catalog families include Whisper, Parakeet, Canary, Moonshine, SenseVoice, and GigaAM.
A native inference worker and supported ONNX engines provide local recognition; backend support depends on the model and hardware.
The Models screen handles selection and downloads. Start with the bundled English model before trying alternatives.

Optional Ollama text cleanup is a separate step, not the speech model. Cloud cleanup requires explicit configuration.

Shared outputs, useful platform extras

Modes

Both platforms · Paste

Frontmost app

Dictate directly where the cursor is.

Both platforms · Copy

Clipboard first

Use the transcript even when direct paste is unavailable.

macOS · Preview

Review text

Inspect the result before sending it on.

macOS · File

Markdown output

Transcribe audio files into local notes.

Windows · History

Revisit a transcript

Review saved transcripts and recordings, or copy the last result from the tray.

Windows · Optional

Text cleanup

Keep recognized text as the starting point. Optional cleanup is separate from the speech model and is off by default.

The differences that matter

Key facts

VachaVox on macOS and Windows
Detail	macOS	Windows
Platform	macOS 14+ · menu-bar app	Windows x64 · tray app with WebView2 settings
Speech runtime	Local Core ML · FluidAudio / Parakeet and WhisperKit	Rust-managed native inference worker and supported in-process ONNX engines
First model	Install a compatible Parakeet or Whisper Core ML model folder; models are not bundled.	Whisper Base (English) is bundled with the preview installer.
Trigger	Hold Fn or a custom hotkey to dictate.	Ctrl + Space; hold to dictate, or tap to start and stop in hold-or-toggle mode.
Outputs	Paste, Copy, Preview, and audio-file transcription to Markdown.	Paste into the active app, copy to the clipboard, and review transcript History.
Saved work	Audio to Markdown · ~/vachavox/output. No automatic transcript History.	Local SQLite History and recording WAV files, managed in the app's profile.
Text cleanup	The documented workflow is local speech transcription and output delivery.	Optional and off by default. Local Ollama is the default target; cloud cleanup requires explicit configuration.
Release	v0.6.3 · Build 23 · ZIP app download	v0.1.2 · unsigned EXE preview · automatic updates disabled
License	MIT + Commons Clause; see the source license for commercial-use terms.	MIT with retained dependency and model notices. Windows source is currently private.

Shared flow, separate desktop implementations

Diagrams

flowchart LR
  A["Shortcut + microphone"] --> B["Local speech model"]
  B --> C["Recognized text"]
  C --> D["Paste or copy into everyday apps"]

The shared core: trigger, transcribe locally, and send text to where you are already working. Model runtimes and desktop integrations differ by platform.

From download to everyday use

Quick start

macOS

macOS 14+ and a compatible local speech model.

Download the macOS ZIP and open VachaVox.
Install a compatible model folder; in Settings → Models, select and load it.
Grant Microphone access. Grant Accessibility access if you want automatic Paste; Copy and Preview work without it.
Focus the destination app, hold Fn or your custom shortcut, speak, and release.
Choose Paste, Copy, or Preview. For an audio file, use file transcription to create a local Markdown document.

Windows

Windows x64. The installer includes the verified English model.

Download and run the Windows setup EXE. It is an unsigned preview.
Choose Whisper Base (English) during onboarding, then select an active microphone in Settings.
Focus the destination app. Hold Ctrl + Space, wait for recording readiness, speak, and release; a tap can start or stop recording.
Review the result in History, or copy the last transcript from the tray menu.
Leave optional cleanup off to start. Configure a local Ollama provider if you want text cleanup.

Source & development

macOS

Swift · Core ML · FluidAudio · WhisperKit

git clone https://github.com/anujraja/VachaVox.git
cd VachaVox

swift build
swift test
Scripts/package_app.sh

Requires the Swift/Xcode toolchain. Models are installed separately; use the public model-installation guide for compatible folders.

Windows

Rust · Tauri 2 · React / TypeScript · native speech inference

# From the Windows source checkout
./scripts/Build-Windows.ps1

Windows source access is required. Maintainer builds use Rust, Bun, Visual C++ Build Tools, Windows SDK, CMake/Ninja, and Vulkan SDK. The public installer runs without this development setup.