This is all true. The one problem I think they really should change though, is they KEEP the voice data they collect by mistake. When the local device thinks it heard the wake word, but the cloud analysis decided it was not said.
There is ZERO reason to keep that data long term. If they want to analyze it to reduce the incidences of this, MAYBE but then it should work like this:
1: Instantly anonymized
2: Sent into an analysis queue
3: Deleted as soon as analysis is complete (with a pretty short window for mandatory deletion).