Alexa Smith
To me, the easiest way to signify your device should start to record is to use two sequential words e.g., "Alexa Smith". There is a 2 second ring buffer which is listening out for the first word. A microphone continually starts and resets that buffer until the word "Alexa" is used in its entirety. This triggers a flag to say that "Smith" needs to be captured using the same buffer within the next 2 seconds otherwise it reverts to quiescent state.
Full-blown recording ensues until a one-word stop fragment is picked up, which saves the recording and puts the device back into quiescent mode. With this scheme the maximum unsolicited recording witll be 2 seconds long.
If coding vulnerabilities arise in coding this simple utility it will bring a whole new meaning to the term "buffer overflow".