This paper discusses two new features for speech activity detection(SAD), using a multi-layer perceptron(mlp) trained to predict phoneme from acoustic features. The first feature is based on the difference between speech and noise histogram of certain phonemes. A scoring mechanism is formulated to score the softmax probabilities of the frames of a phoneme. The second feature is based on the correlation between softmax probabilities of the edge frames for certain phoneme transitions. A probabilistic approach is formulated to score the phoneme transition. Relevant datasets are used to prove the robustness of the proposed features in terms of speech activity detection.