Robust signal/noise discrimination for wideband speech and audio coding

We present a robust discrimination method to separate information carrying signals from ambient noise for wideband coding applications. The method consists of a two-stage procedure. First, a local decision is made based on a set of extracted parameters to update the estimated noise level. The parameters have been chosen to reliably detect speech as well as music signals. Then, the final decision is made based only on a frequency dependent signal to noise ratio (SNR). The noise level update does not depend on the final decision to prevent the discriminator from locking when noise level changes suddenly. The performance is compared with the performance of the voice activity detector (VAD) of G.729, Annex G.