Hands-free voice activation of personal communication devices

We propose a real-time speaker-independent voice activation system composed of an HMM-based continuous speech recognition algorithm and a post-processing component. The post-processing component validates the recognition results by applying a set of criteria and confidence measures. Voice activation in this context refers to a system designed to continuously listen to incoming unconstrained speech and to recognize a pre-defined activation command while rejecting any other extraneous speech or noise. The proposed system is intended for turning a device on or for activating a main speech recognition application upon recognizing the activation utterance. This paper includes a description of the overall system, the postprocessing algorithm formulation, and a discussion of the final results. The proposed system has been trained and tested on a database comprised of 2000 North American English speakers. The system yields a 0.55% false alarm rate, and a 15% false rejection rate for both isolated and embedded utterances.