Voice activity detection using AdaBoost with multi-frame information

A noise robust scheme for voice activity detection (VAD) that employs a combination of both intra- and inter-frame acoustic features is presented in this paper. As intra-frame features full-band energy and mel-frequency cepstrum coefficient (MFCC) are calculated whereas integrated bispectrum is estimated as inter-frame features. The parameters combined by intra- and inter-frame features are sorted out by using adaptive boosting (AdaBoost) algorithm, thereby resulting in a better performance in contrast to a scheme with only a single feature extracted from every frame. On the basis of VAD evaluation framework, CENSREC-1-C (Corpora and Environments for Noisy Speech RECognition-1 Concatenated), the accuracy of the proposed VAD scheme is examined. The results of numerical experiments suggest that the performance of the proposed VAD scheme significantly outperforms conventional methods in real noisy environments.