I was surprised they store a real recording. I thought they submitted some pre-processed form which saves bandwith, partly anonymizes the speech and then hands it off for recognition.