Looks like it can all be done offline + OSS except for Speech-To-Text. But I think the key idea is even if it's using cloud for processing, you can at least trust it isn't listening when it shouldn't be (as long as always-on, wake word detection is offline).