论文部分内容阅读
针对欠定混合语音信号模型,提出一种基于独立分量分析和二值掩膜相结合的语音分离新算法.首先,由并列放置的两个全指向性麦克风采集混合声音信号,通过一阶差分麦克风阵列技术使得两路混合信号满足瞬时混合模型;然后,应用两输入两输出独立分量分析方法对两路信号进行分离并估计分离信号的二值掩膜,循环迭代进而提取目标语音信号;最后,合并同源语音信号,增强各独立声源.仿真结果表明:该方法不仅适用于瞬时混合模型,对卷积混合模型同样有效;在瞬时混合条件下信噪比增益平均达到12.41dB,在卷积混合条件下信噪比增益平均为5.28dB.试验结果表明:算法在半消音室环境下能准确分离来自不同方位的三个声源,提取的目标语音都具有较高的清晰度与可懂性.
Aiming at underdetermined mixed speech signal model, a new speech separation algorithm based on independent component analysis and binary mask is proposed.Firstly, two omni-directional omni-directional microphones are used to collect mixed acoustic signals, The array technique makes the two mixed signals satisfy the instantaneous mixed model. Then, two-input and two-output independent component analysis methods are used to separate the two signals and estimate the binary mask of the separated signal. The iterative process is performed to iteratively extract the target speech signal. Finally, The simulation results show that the proposed method is not only suitable for instantaneous mixed model, but also effective for convolutional mixed model. Under the condition of instantaneous mixing, the average gain of signal-to-noise ratio reaches 12.41dB, The average gain of signal-to-noise ratio under the condition is 5.28dB.The experimental results show that the algorithm can accurately separate the three sound sources from different azimuths in the semi-anechoic chamber environment, and the extracted target speech has higher definition and intelligibility.