【摘要】 提出了一种用神经网络模型和信号子空间特征分解相结合进行语音增强的方法，该方法利用了神经网络并行处理、高速计算的能力和语音信号的短时平稳的特性．它既克服了传统的谱相减法中残留“音乐噪声”的缺陷，又可以实时、有效地增强语音，能够满足一些语音处理与识别系统的需要．更多还原

【Abstract】 This paper presents a speech enhancement approach based on signal subspace decomposition, which is implemented in real time by the use of neural networks. This approach takes advantages of the parallel processing and high speed computational capability of neural networks. It can amend a residual musical noise in traditional spectral subtraction approach. This approach can also be applied to some other fields, such as image processing and mobile channel equalization.更多还原