Abstract: Based ondeep neural network (DNN), a pitch detection algorithm is proposed to solve the challenging problem of pitch detecting in strong noisy background.In this paper, we estimate pitch using supervised learning, where the probabilistic pitch states are directly learned from noisy speech data. The feedforward deep neural network is trained on stastic frame-level acoustic features to produce accurate probabilistic outputs of pitch states, which are then connected into pitch contours by Viterbi decoding.The experimental results show that the proposed algorithm are robust to different noise conditions and can even be applied to reverberant speech.The proposed approach outperforms other state-of-art pitch tracking algorithms.