RESEARCH ON ENGLISH SPEECH ENHANCEMENT ALGORITHM BASED ON IMPROVED SPECTRAL SUBTRACTION AND DEEP NEURAL NETWORK

In order to solve the introduced unstructured voiceless problems of conventional spectrum subtraction in English speech signals enhancement, this paper proposes a novel English speech signals enhancement algorithm. This algorithm uses an improved minimal controlled recursive averaging (IMCRA) method to estimate noise spectrum, and tracks the estimated noise spectrum in real time. Then, the deep neural network (DNN) is used to construct the nonlinear mapping function of log amplitude spectrum between speech with noises and ideal pure speech for English speech enhancement. To validate the feasibility and effectiveness of the proposed algorithm, the standard IEEE speech signals and Noise-91 noise signals are used for experiments. Experimental results have shown that the proposed IMCRA method has stronger ability to avoid noises in speech signals, and the DNN method can well recover the speech components and spectrum structure polluted by noises. To enhance English speech in daily international speech communication, the proposed combination method has strong robustness to various real noise environments, and brings significant improvement to interpersonal communication and human computer communication.