The model downloads once, then stays cached. Speech recognition runs on your device, so your audio never leaves it. The download size depends on the model above.
Works in Chrome, Firefox and Safari. Longer recordings take a while on a phone.