[acpetrov]

> People perhaps assume all audio is processed locally and/or isn't stored except ephemerally.

Is there info on this from Amazon? I'd assume they do what you mention, have the speech recognition in the Alexa and just send up the words, not the entire audio stream.

Interesting implication of that in this case: if the actual audio is gone and you just have the transcription, is that valid as evidence? Or is it something closer to hearsay

[detaro]

It's cloud based. The device recognizes the wake-word offline and starts streaming to the recognition service. You can access a log of recordings through the app, so they are kept at least for a while.