An end-to-end single channel speech separation algorithm based on deep acoustic feature is proposed. The traditional acoustic feature extraction methods require the Fourier transform, discrete cosine transform and other operations. This will cause speech energy loss and long latency. In order to improve these problems, the original waveform of the speech signal is used as an input to a deep neural network, deeper acoustic features of the speech signal are learned through a network model. Objective evaluation shows that the proposed algorithm not only improves the performance of speech separation effectively, but also reduces the time delay of speech separation algorithm.