A voice activation system is described including a first voice activity detector for receiving a first signal from a microphone; a second voice activity detector for receiving a second signal from a speaker; a voice detector output coupled to the output of the first voice activity detector and the second voice activity detector. The first detector and the second detector are operable to generate an output signal in response to a candidate speech signal received on the respective detector inputs and the voice activation system is configured to generate a voice detector output signal when at least one of the first detector output and the second detector output indicate that speech has been detected. The voice activation system may reduce the average power consumption and increase the recognition rate of an always-on voice activation solution for headsets or other mobile audio devices.