Most products use local low-power always-on listening, and only after recognizing a specific wake word do they upload the subsequent speech, in order to reduce power consumption and lessen privacy risks.