Local Models

100% offline transcription with Whisper

Run OpenAI's Whisper model directly on your device for complete privacy and offline use.

Available Models

English-Only Models

  • base.en: Fastest, lower accuracy
  • small.en: Good balance of speed and accuracy
  • medium.en: High accuracy, slower
  • large-v3-turbo-en: Best accuracy for English

Multilingual Models

  • base: 100+ languages, fastest
  • small: 100+ languages, good balance
  • medium: 100+ languages, high accuracy
  • large-v3: Best accuracy, slowest
  • large-v3-turbo: Near-best accuracy, faster

GPU Acceleration

Speakly automatically uses GPU acceleration when available:

  • macOS: Metal (Apple Silicon M1/M2/M3)
  • Windows/Linux: Vulkan

Model Management

Download, switch, and delete models in Settings > Transcription.

Recommended
For most users, we recommend starting with 'small.en' (English) or 'small' (multilingual) for a good balance of speed and accuracy.
Documentation - Speakly