An efficient VAD algorithm based on constant False Acceptance rate for highly noisy environments
Charaf Eddine Chelloug, Atef Farrouki · 2016
Voice activity detection (VAD) is frequently used in many speech processing applications to distinguish the active voice regions from the silent intervals. In speech compression systems, the VAD represents an important build-in component to reduce the network bandwidth consumption through the discontinuous transmission mode. In this paper, we propose a VAD scheme based on adaptive threshold to maintain a prescribed false acceptance rate. Full-band energy-based sequential tests have been implemented in order to discard, or to accept the frame under investigation as active voice region. The most attractive feature of the proposed algorithm consists of its ability to dynamically update the noise level estimator according to the current ambient environment. By assuming a long term stationary nature of the speech signal, we also developed a smoothing procedure which analyzes the last partial decisions, in order to generate the final VAD about the frame under test. The performances of the proposed approach are evaluated and compared to the standard G.729B VAD in several situations, including various acoustical environmental noises with different SNR. The results analysis has been carried out by using the experimental NOIZEUS database as well as realistic recorded speech signals.