Dictaitor

On-device speech-to-text iOS app powered by NVIDIA Parakeet TDT 0.6B v3 running on Apple Neural Engine.

How it works

1. Tap the mic button and speak
2. Text appears in real-time as you talk (via Apple Speech framework)
3. Tap stop — Parakeet v3 produces the final high-accuracy transcription on-device
4. Result is automatically copied to clipboard (syncs to Mac via Universal Clipboard)

Features

• Real-time feedback — live transcription displayed as you speak
• High-accuracy final result — Parakeet v3 (0.6B parameters) runs on Apple Neural Engine
• 25 European languages — Polish, English, German, French, Spanish, and more
• Fully offline — after initial model download (~600MB), no network needed
• Auto-copy — transcript copied to clipboard on stop
• Universal Clipboard — paste directly on your Mac with ⌘V

Requirements

• iPhone with Apple Silicon (A14+) — iPhone 12 or newer
• iOS 17.0+
• ~700MB free storage for model download on first launch
• WiFi required for initial model download only

Building

Prerequisites

• Xcode 16+ with iOS SDK
• Apple Developer account (free or paid)

Build & Run

# Point to your Xcode installation
sudo xcode-select -s /path/to/Xcode.app/Contents/Developer

# Build
xcodebuild -project Dictator/Dictator.xcodeproj \
  -target Dictator -sdk iphoneos -arch arm64 \
  -allowProvisioningUpdates \
  DEVELOPMENT_TEAM=YOUR_TEAM_ID \
  build

# Install on connected device
xcrun devicectl device install app \
  --device YOUR_DEVICE_ID \
  Dictator/build/Release-iphoneos/Dictator.app

Or open Dictator/Dictator.xcodeproj in Xcode, select your device, and hit Run.

Finding your device ID

xcrun devicectl list devices

Architecture

┌─────────────────────────────────────────────────┐
│                  Dictaitor App                   │
├─────────────────────────────────────────────────┤
│                                                 │
│  ┌─────────────┐     ┌──────────────────────┐  │
│  │ AVAudioEngine│────▶│ Apple SFSpeechRecognizer│
│  │  (16kHz mono)│     │  (live partial results) │
│  └──────┬───────┘     └──────────────────────┘  │
│         │                                       │
│         ▼ on stop                               │
│  ┌──────────────────┐                           │
│  │   Parakeet v3    │                           │
│  │  (CoreML / ANE)  │──▶ Final text ──▶ Clipboard
│  │  via FluidAudio  │                           │
│  └──────────────────┘                           │
│                                                 │
└─────────────────────────────────────────────────┘

Credits

• FluidAudio — Swift framework for on-device audio ML
• NVIDIA NeMo Parakeet — ASR model
• FluidInference CoreML conversion — CoreML model artifacts

License

MIT — see LICENSE