Robust Voice Activity Detection Against Non Homogeneous Noisy Environments

In this work, an improved voice activity detection (VAD) algorithm is presented to deal with non stationary noisy environments. The proposed approach is based on adaptive thresholding to regulate the False Acceptance (FA) ratio in absence of active voice. Sequential hypothesis tests, using full band energy, have been carrying out to reject or to classify the frame under processing as a voiced segment. The main advantage of the proposed technique consists of its capability to automatically update the level of background noise, by taking into account the current environment. Performances and real time behavior have been analyzed and compared to the modern standard G.729-B, by implementing the proposed VAD architecture on a Micro Controller Unit-based processing system.