ABSTRACT
This chapter describes a software-based speech activity detector that is lightweight, fast, and relatively inexpensive in terms of CPU cycles. In any packet-based voice transmission system, such a detector can significantly impact bandwidth utilization by making it possible to transmit only when the speaker is actually talking. Unlike circuit-switched systems, in which bidirectional channel bandwidth is dedicated throughout the life of the channel despite the intermittent nature of speech transmission, in theory, a packet system only uses bandwidth for the information that it has to transfer, and only in the direction of the transfer. In order to realize that gain in a speech context it is necessary to discriminate accurately and quickly between periods of speech activity and periods of relative silence. An important part of the research project that led to the development of the STNi was proving that the touted efficiency gains were practically attainable. Speech Activity Detection, therefore, was an integral part of the STNi design.
