WalkWrite - Voice Notes with On-Device AI

WalkWrite is a voice note app that uses on-device Whisper 3.5 Turbo and Gwen 3 (but you could easily swap those models) to transcribe and enhance recordings on device. All processing happens locally on the device - no internet connection required, no data leaves the phone.

This open-source version serves as a template and example for iOS developers who want to integrate local AI models (Whisper & LLMs) into their own apps. The full app is available on the App Store.

Note: This open-source version has no usage limits - all features are available without restrictions. The optional one time in-app purchase remains for users who wish to support development and updates.

Features

• 🎙️ High-quality voice recording with real-time audio that is safely always saved on the users device.
• 🤖 On-device transcription using OpenAI's Whisper (large-v3-turbo model)
• ✨ AI-powered enhancements using Qwen-3 0.6B LLM:
  • Grammar and punctuation correction
  • Filler word removal
  • Automatic summarization
  • Key idea extraction
• 🔒 Complete privacy - all processing happens locally
• 📱 Native iOS app built with SwiftUI
• 🎯 Offline first - works offline on an airplane or in the woods

Screenshots

Requirements

• iOS 17.0+
• iPhone with A12 Bionic chip or newer (for Neural Engine)
• ~3GB free storage for models, App ships with the models but these could be easily made optional downloads and it could ship with smaller models
• Xcode 15.0+

Building from Source

1. Clone the repository:

git clone https://github.com/lbacaj/WalkWrite-opensource.git
cd WalkWrite-opensource

2. Install Git LFS and pull the model files:

brew install git-lfs
git lfs install
git lfs pull

Alternatively, you can download the models directly from Hugging Face if you prefer.

3. Build the whisper.cpp framework:

./build-whisper-xcframework.sh

4. Configure your development environment:

./setup-developer.sh

This will set up your development team and bundle identifier.

5. Open the project in Xcode:

open WalkWrite.xcodeproj

6. Let Xcode resolve Swift Package dependencies (this may take a few minutes)

7. Build and run on your device (simulator not recommended due to model performance)

Architecture

WalkWrite uses a Swift architecture with:

• SwiftUI for the UI layer
• Swift Actors for thread-safe model management
• whisper.cpp for speech-to-text transcription
• MLX Swift for running the Qwen LLM
• Core ML for optimized on-device inference

Key Components

• WhisperEngine.swift - Manages whisper.cpp transcription
• LLMEngine.swift - Handles Qwen model inference via MLX
• RecorderViewModel.swift - Recording logic and audio processing
• PostProcessor.swift - Orchestrates the enhancement pipeline
• NoteStore.swift - Persistence and data management

Models

The app includes two AI models:

1. Whisper Large-v3-turbo (Q5_0 quantized, ~574MB)

   • High-quality speech recognition
   • Optimized for mobile devices

2. Qwen-3 0.6B (~1.4GB)

   • Lightweight language model
   • Grammar correction and text enhancement

Privacy

WalkWrite is designed with privacy as the top priority:

• ✅ All AI processing happens on-device
• ✅ No internet connection required for core features
• ✅ No telemetry or analytics
• ✅ Audio files stored locally in app sandbox
• ✅ Export only when explicitly requested by user

Using as a Reference

This project is designed as a reference implementation for developers who want to:

• See how to integrate whisper.cpp in an iOS app
• Learn how to run local LLMs with MLX Swift
• Build privacy-focused AI features without servers

Feel free to use any code from this project in your own apps!

License

This project is licensed under the MIT License - see the LICENSE file for details.

Use as a Template

This project is designed to help iOS developers learn how to:

• Integrate whisper.cpp for on-device speech recognition
• Run small LLMs locally using MLX Swift
• Manage memory efficiently when running multiple AI models
• Build privacy-first AI features without server dependencies
• Handle background processing for long-running AI tasks

Feel free to use this code as a starting point for your own AI-powered iOS apps!

Acknowledgments

• whisper.cpp by Georgi Gerganov
• MLX Swift by Apple
• Qwen by Alibaba Cloud
• Not fully "vibe coded" but AI did help build this, so it deserves acknowledgment

Author

Created by Louie Bacaj