Alexa devices use an onboard DSP to detect the wakeword and maintain a rolling audio buffer. On a positive match, the DSP wakes the main CPU which combines the saved buffer and any following speech and uploads it to the cloud where Alexa lives so she can try to figure out what you meant.
No audio is uploaded without being triggered by a wakeword. Also, the "mute" button physically cuts power to the mic, and the indicator LED is hardwired to the power rail as a failsafe indicator.
Not going to get much out of me then, most of what Alexander hears is what's on the TV or music I listen to. If they want to train alexa on that, their fucked