The purpose of this study was to explore voice activity detection (VAD) in a subject with implanted electrocorticographic (ECoG) electrodes. Accurate VAD is an important preliminary step before decoding and reconstructing speech from ECoG. For this study we used ECoG signals recorded while a subject performed a picture naming task. We extracted time-domain features from the raw ECoG and spectral features from the ECoG high gamma band (70-110Hz). The RelieF algorithm was used for selecting a subset of features to use with seven machine learning algorithms for classification. With this approach we were able to detect voice activity from ECoG signals, achieving a high accuracy using the 100 best features from all electrodes (96%) or only 12 features from the two best electrodes (94%) using the support vector machines or a linear regression classifier. These findings may contribute to the development of ECoG-based brain machine interface (BMI) systems for rehabilitating individuals with communication impairments.