VoiceScribe

Native local AI dictation for macOS, powered by Qwen3-ASR and MLX.

Press Option + Space, speak, press it again, and VoiceScribe pastes clean text back into the app you are using.

VoiceScribe is built for people who want a real offline speech-to-text app for Mac, not a desktop shell hiding a Python daemon in the background.

Onboarding

VoiceScribe now opens with a more polished first-run flow designed to feel closer to a native macOS product:

• a cinematic onboarding window that previews the HUD in context
• local model selection directly in the flow
• a simple shortcut handoff before the first real dictation
• local preparation explained before the app starts working in the background

Speak anywhere • Choose your local model • Ready from a single shortcut

The first launch can take longer because the selected Qwen3-ASR model is downloaded once and cached locally on your Mac.

Dictation Flow

Once onboarding is complete, the app stays intentionally simple:

1. Press Option + Space
2. Speak normally
3. Press Option + Space again
4. VoiceScribe transcribes locally and pastes the result back into your current app

Ready	Recording

Highlights

• Fully local speech-to-text for macOS on Apple Silicon
• Native Swift 6 + MLX runtime with Qwen3-ASR
• No cloud dependency, no Python daemon, no subprocess bridge
• One hotkey, one floating HUD, one fast dictation loop
• Automatic clipboard copy and paste injection
• Strong English and French dictation

Why Native MLX

VoiceScribe uses a pure Swift + MLX pipeline for offline dictation on macOS.

That gives you:

• fewer moving parts
• better crash isolation
• simpler packaging
• direct Apple Silicon acceleration through Metal
• one native concurrency model across UI, audio, and inference

Why Qwen3-ASR

Qwen3-ASR is the core reason VoiceScribe feels competitive as a local dictation app on Apple Silicon.

• fast enough for real short-form dictation workflows
• strong quality-to-latency balance in local use
• better multilingual behavior for English and French
• fewer weird artifacts than many offline speech-to-text stacks
• a good fit for MLX + Metal on Mac

Supported Models

VoiceScribe supports mlx-community/Qwen3-ASR variants only.

Default model:

• mlx-community/Qwen3-ASR-1.7B-8bit

Install

Requirements:

• macOS 14+
• Apple Silicon
• microphone permission
• accessibility permission for automatic paste

Latest release:

• Download VoiceScribe

Build from source:

git clone https://github.com/Flovflo/VoiceScribe.git
cd VoiceScribe
swift build -c release --arch arm64
./package_app.sh
open VoiceScribe.app

Validation

Fast suite:

swift test

Optional MLX validation:

VOICESCRIBE_RUN_MLX_TESTS=1 swift test --filter AudioFeatureTests

Optional real ASR validation:

VOICESCRIBE_RUN_ASR_TESTS=1 swift test --filter NativeEngineTests

Docs

• Native MLX notes
• AGENTS.md

License

MIT