面向旋翼飞机螺旋桨干扰的AM通信语音信号智能增强方法

为改善旋翼飞机空地语音通信质量，针对旋翼飞机螺旋桨造成的幅度调制(Amplitude Modulation,AM)信号复杂多频干扰以及恶劣机舱背景噪声，提出了一种通信语音时频掩膜智能增强方法，从而实现对机舱噪声与复杂干扰的有效抑制。该方法首先对原始时域语音信号进行分帧与加窗，通过短时傅里叶变换获取幅度谱与相位谱；然后将原始幅度谱作为网络输入，采用深度神经网络分析其语音信号的特征，采用长短期记忆网络挖掘语音信号的时序上下文信息，实现对语音时频掩膜的准确估计，并将其用于增强原始幅度谱以得到网络输出；最后结合原始相位谱，通过逆短时傅里叶变换获得增强后的时域语音信号。仿真与实际测试表明，该方法可有效抑制旋翼飞机环境下的干扰噪声，提高通信语音信号质量。

Deep speech enhancement based on time-frequency mask for propeller interference of rotor aircraft

In order to improve the quality of speech communication in air-to-ground rotor aircraft,a speech enhancement method is proposed to deal with complex multi-frequency interference caused by the propeller and the harsh noise in the cabin.The method is based on deep neural network,which estimates the time-frequency(T-F) mask and suppresses the noise and the interference effectively.The original time-domain speech signals are framed and windowed first,and the amplitude and phase spectrum are obtained through short-time Fourier transform(STFT).Then the original amplitude spectrum is taken as the input of the network,which analyzes the features of the speech signal.The main hidden layer,the Long Short-Term Memory(LSTM) network mines the sequential information in the speech signal,realizing the accurate estimation for the T-F mask of speech signal.The T-F mask augments the original amplitude spectrum,which is the output of the network.Combined with the original phase spectrum,the enhanced time-domain speech signal is reconstructed by inverse STFT(ISTFT).Simulation and test show that the proposed method can suppress the interference noise in the rotor aircraft environment effectively and improves the quality of the communication voice signal.