LLMVoice

Real-time voice transcription and AI summarization for iOS/macOS.

Overview

LLMVoice combines Apple's latest speech recognition APIs with on-device LLM inference to transcribe and summarize voice recordings in real-time. Built for iOS 26+ using Swift 6, it runs entirely on-device with no network calls.

Features

• Real-time Transcription: Progressive speech-to-text using iOS 26's SpeechTranscriber
• Multi-language Support: Transcribe in English, Spanish, French, German, Italian, Portuguese, and more
• On-device LLM: Summarize transcriptions using MLX-powered models (Qwen, Gemma, Llama)
• Streaming Generation: Watch AI summaries appear token-by-token with live metrics
• Apple Intelligence Fallback: Uses Apple's Writing Tools API when available
• Model Management: Download and manage multiple LLM models
• Performance Metrics: Track tokens/second, generation time, and model performance

Requirements

• iOS 26.0+ / macOS 15.0+
• Xcode 16+
• Swift 6
• Device with A12 Bionic or newer for MLX models
• Device with Apple Neural Engine for Apple Intelligence

Tech Stack

• Speech Recognition: SpeechTranscriber, SpeechAnalyzer (iOS 26 SDK)
• Audio Processing: AVAudioEngine, AVFoundation
• LLM Inference: MLX Swift
• Model Loading: swift-transformers
• UI: SwiftUI with Observable macro
• Architecture: MVVM with MainActor isolation

Building

1. Clone the repository:

   git clone https://github.com/AndreFrelicot/llmvoice.git
   cd LLMVoice

2. Open in Xcode:

   open LLMVoice.xcodeproj

3. Select your target device/simulator (iOS 26+)

4. Build and run (⌘R)

Project Structure

LLMVoice/
├── Managers/
│   ├── AudioManager.swift              # Audio capture
│   ├── TranscriptionManager.swift      # Speech-to-text
│   ├── SummarizationManager.swift      # AI summarization
│   └── MLXSummarizationManager.swift   # MLX model inference
├── ViewModels/
│   └── RecordingViewModel.swift        # Main app state
├── Views/
│   ├── ContentView.swift               # Main UI
│   ├── SummariesListView.swift         # Results list
│   └── ModelPickerView.swift           # Model selection
└── Models/
    ├── MLXModel.swift                  # Model definitions
    └── Summary.swift                   # Data models

Supported Models

• Qwen 3 (0.6B) - 335MB, 100+ languages, 32k context - Recommended
• Qwen 2.5 (0.5B) - 150MB, 29+ languages, 32k context
• Gemma 3 (1B) - 300MB, 140+ languages, 32k context
• Gemma 3 QAT (1B) - 733MB, 140+ languages, quantization-aware trained 4-bit model
• Llama 3.2 (1B) - 500MB, 8 languages
• Custom MLX-compatible models via Hugging Face

License

MIT License - see LICENSE file for details

Attribution

Built with:

• MLX Swift by Apple ML Research
• swift-transformers by Hugging Face
• Apple's SpeechTranscriber API (iOS 26 SDK)

---

Note: This app requires iOS 26 beta SDK. Speech transcription and on-device AI features require compatible hardware.