Don't type,
just speak

The open-source voice-to-text app that turns speech into clean, polished text in every macOS app. Fully offline. No cloud. No subscriptions.

Download Now View on GitHub

macOS 14+ · Apple Silicon · Apache 2.0

VibeFlow
Listening...
🎙 Capture
→
📝 Transcribe
→
✨ Polish
→
📋 Paste
VS Code
Slack
Notion
Terminal
Safari
Xcode

See It In Action

Watch the demo

Hold a key, speak, release — see how VibeFlow turns voice into polished text in any macOS app.

Why VibeFlow

Voice that respects
your privacy

Unlike cloud-based dictation tools, VibeFlow runs entirely on your Mac. Your voice never leaves your device.

VibeFlow

Open Source
  • ✓ 100% local processing
  • ✓ No account required
  • ✓ No subscription fees
  • ✓ Works without internet
  • ✓ Apache 2.0 licensed
  • ✓ Extend with your own engines

Cloud Dictation

Proprietary
  • ✗ Voice sent to servers
  • ✗ Account + login required
  • ✗ Monthly subscription
  • ✗ Requires internet
  • ✗ Closed source
  • ✗ Limited customization
🎯

Universal Dictation

Works across any macOS app — Slack, VS Code, Notion, browsers, terminals. Hold a hotkey, speak, release. Done.

🧠

Dual Speech Engines

Apple Speech Recognition for speed, or WhisperKit for offline Whisper-quality transcription via Neural Engine.

✨

AI Text Cleanup

Local Qwen 0.5B via MLX polishes your speech — removes filler words, fixes grammar, formats text. All on-device.

✈️

Fully Offline

WhisperKit + Local SLM = zero network dependency. Dictate on a plane, in a tunnel, anywhere.

📖

Custom Dictionary

Add technical terms — Kubernetes, Terraform, gRPC — for accurate recognition of your domain vocabulary.

🎨

Writing Styles

Casual, Professional, Creative, or Technical. VibeFlow adapts its text cleanup to your preferred tone.

How It Works

Three seconds from
thought to text

1
fn

Hold the hotkey

Press and hold Fn (or your custom key). The Dynamic Island HUD appears with a live waveform.

2

Speak naturally

Talk normally. Say "um" and "like" all you want — the filler removal pipeline strips them out.

3

Release & paste

Let go. VibeFlow transcribes, cleans, and pastes polished text into whatever app you're using.

Hotkey
Speech Engine
Apple Speech / WhisperKit
Filler Remover
Regex patterns
Text Processor
Qwen 0.5B / Remote LLM
Paste

Engine Options

Mix and match
your pipeline

Two stages, two choices each. Pick what fits your workflow — switch anytime in Settings.

Stage 1: Speech-to-Text

Built-in

Apple Speech

Uses macOS SFSpeechRecognizer. Fast, zero setup, supports contextualStrings for custom dictionary terms.

0 MB download Instant start
Offline

WhisperKit

OpenAI Whisper running on Apple Neural Engine via CoreML. Best accuracy, fully offline, auto-downloads models.

~80 MB model ~100 MB RAM

Stage 2: Text Cleanup

Offline

Local SLM

Qwen 0.5B quantized to 4-bit, running on Apple GPU via MLX Swift. Polishes text without any network call.

~350 MB model ~500 MB RAM
Flexible

Remote LLM

Any OpenAI-compatible endpoint — GPT-4, Claude, local Ollama. Use the LiteLLM client for maximum flexibility.

0 MB local Requires network

Resource Usage

ComponentDiskRAM
App binary~20 MB~50 MB
WhisperKit (base)~80 MB~100 MB
Qwen 0.5B (4-bit)~350 MB~500 MB
Total (fully offline)~450 MB~650 MB

Runs comfortably on a base MacBook Air M1 with 8 GB RAM.

Architecture

Built for
extensibility

Protocol-based engine system lets you hot-swap speech and text processing at runtime. Add your own engines by conforming to a protocol.

SpeechRecognitionService

protocol SpeechRecognitionService
  • AppleSpeechEngine
  • WhisperEngine
  • + your own

TextProcessingService

protocol TextProcessingService
  • LocalSLMProcessor
  • RemoteLLMProcessor
  • + your own

Core Services

ConversationController
  • FillerRemover
  • ModelDownloadManager
  • PermissionsHelper

Technologies

SwiftUI SwiftData AVFoundation Speech.framework WhisperKit MLX Swift AppKit CoreAudio

Changelog

What's new

VibeFlow ships fast. Here's what landed recently.

v1.2 Latest Mar 28, 2026

Dark Mode & UX Polish

  • Dark Mode — Appearance toggle (Light / Dark / System) in Settings, saved across launches. All views render correctly in both modes.
  • Richer History — Tap any transcription to see the full text, WPM speed, and duration. Average WPM now shown on Dashboard. WPM shown inline on each history row.
  • Recording Quality — 300 ms grace delay before stopping so the last word is never cut off.
  • Settings Cleanup — Microphone permissions surfaced at the top, redundant Formality setting removed.
  • Bug Fixes — Window remembers its size between launches. AI model used is now accurately logged per transcription.
Full release notes →
v1.1.0 Mar 28, 2026

Memory & Reliability

  • Critical Memory Fix — Eliminated a staircase memory leak (400 MB → 4+ GB) caused by 7 reactive observers rebuilding models on every keystroke. Models now only rebuild on Save Settings.
  • Model Load Error Surfacing — If WhisperKit or the local SLM fails to load, a red banner with Retry / Change Model appears on the Dashboard. Recording is blocked until the model is healthy.
  • HUD Processing State — Spinner + "Processing…" shown after release while transcription + AI cleanup runs. A red error pill appears for 3 s on failure.
  • Settings Save UX — All changes are held locally until you press Save Settings. Navigating away with unsaved changes shows a discard confirmation.
Full release notes →
v1.0.0 Mar 24, 2026

Initial Release

  • Universal dictation across any macOS app via global hotkey
  • Dual speech engines: Apple Speech & WhisperKit
  • Dual text cleanup: Local SLM (Qwen 0.5B) & Remote LLM
  • Dynamic Island HUD with live waveform
  • Custom dictionary, writing styles, transcription history
Full release notes →
View all releases on GitHub →

Start flowing

Open source voice dictation for engineers. 100% local. 100% free.

Download Now View Source
git clone https://github.com/agarwalvivek29/VibeFlow.git && cd VibeFlow && open VibeFlow.xcodeproj