Hacker News Reader

Why can't they just have full on-device transcription models? It's lossy, but extremely compressed.

Because they're cheap, low powered devices.

Even cheap devices can run transcription inference.