Explore AI

Transcribing and speaking

Turning recordings into text, or text into a voice. Small models do this well.

Turning recordings into text, or text into a voice. Small models do this well.

What we would try first

  • Whisper Large v3 Turbo 4 GB at Q4 Transcription that runs comfortably on a modest machine and beats most hosted services for accuracy. The best-value local AI in any category.
  • Kokoro 82M 1 GB at Q4 82 million parameters for speech that sounds better than systems a hundred times its size. Runs on a phone.
  • Chatterbox 3 GB at Q4 Voice cloning with an MIT licence, which is rare — the popular alternatives are non-commercial or restricted.