WhisperKit — Local AI (Level 2)

WhisperKit is Lagos’s recommended transcription engine for users who want the best accuracy without sending audio to a server. It runs OpenAI’s Whisper model locally on your Mac using Apple’s Core ML framework.

How It Works

WhisperKit runs the Whisper neural network entirely on your device’s Neural Engine and GPU. It achieves accuracy comparable to cloud services — while keeping your audio completely private.

Models Available

Lagos supports two WhisperKit models:

Model	Size	WER	Speed
Standard (base)	~150 MB	~5%	Fast
Turbo (large-v3-turbo)	~1.5 GB	~1%	Medium

Start with Standard for a quick download. Upgrade to Turbo for the highest accuracy.

Setting Up WhisperKit

1. In Lagos Settings, select WhisperKit as your engine
2. Click Download Model next to your preferred model
3. Wait for the download to complete (~150 MB for Standard, ~1.5 GB for Turbo)
4. WhisperKit is ready — no further configuration needed

Models are downloaded once and stored at ~/Library/Application Support/Lagos/models/.

Accuracy

• Standard model: ~5% WER on average English speech
• Turbo model: ~1% WER — competitive with commercial cloud services

WhisperKit is particularly strong at technical terminology, proper nouns, and non-standard vocabulary. It also handles accents better than Apple Speech.

Privacy

Audio is never sent to any server. All transcription happens on-device using your Mac’s Neural Engine. Nothing leaves your machine.

Language Support

WhisperKit supports 100+ languages via the Whisper model’s multilingual capabilities. Set your language in Lagos Settings → Language, or use Auto-detect.

Hardware Requirements

WhisperKit works on Apple Silicon (M1 and later). It also runs on Intel Macs but is significantly faster on Apple Silicon.

When to Use WhisperKit

• You want the best accuracy without a cloud subscription
• Privacy is important — no audio should leave your device
• You’re on a reliable internet connection for the one-time model download
• You want full language support