An effective speech brain machine interface requires selecting the best cortical recording sites and signal features for decoding speech production, but also minimal clinical risk for the patient. Motivated by this need to reduce patient risk, the purpose of this study is to detect voice activity (speech onset and offset) automatically from spatial-spectral features of electrocorticographic signals using the optimal number of sensors (minimal invasiveness). ECoG signals were recorded while a subject performed two different syllable repetition tasks. We found that the optimal frequency resolution for detecting voice activity is 8 Hz using 31 sensors out of 55, achieving 98.2% accuracy by employing support vector machines (SVM) as a classifier, and that acceptable accuracy of 96.7% was achieved using 15 sensors, which would permit a less invasive surgery for the placement of electrodes. The proposed voice activity detector may be utilized as a part of an ECoG-based automated natural speech BMI system.