An approach to an optimized voice-activity detector for noisy speech signals

In this paper, we present a voice activity detection (VAD) algorithm for noisy speech which is necessary for many applications such as source coding or speech enhancement methods. The proposed algorithm is characterized by a large conditional detection probability required for noise reduction and shows considerable improvement compared to known methods. The detection procedure is based on the output of an adaptive prediction error filter presented in [2]. Our approach utilizes this prediction error signal to generate a highly reliable voice activity detection with the described procedures. For comparing the results of different methods for voice activity detection, we apply the receiver operating characteristic (ROC). This ROC allows to judge the VAD quantitatively. Additionally, we utilize the ROC to motivate and formulate a new procedure to adapt the VAD threshold automatically to the prevailing SNR which minimizes the required heuristic parameter settings.