基于时频域的具有延迟的欠定盲分离

本文针对衰减-时延混合模型，提出一种改进欠定语音盲分离算法。利用独立语音在时频域上的近似稀疏性，提出势函数面-聚类方法，通过此方法可直接获得源信号的个数和各声源方位信息，省去传统聚类方法中衰减-时延的配对步骤，提出的新势函数具有很好的适应性。分离阶段，利用势函数面聚类方法估计声源信号方位信息重构声源相对传递函数，采用非线性时频掩蔽方法分离感兴趣目标语音信号，实现了不同方位信息声源信号的欠定分离，放宽了对稀疏性的限制。仿真实验证明了此分离方法的可行性和有效性,弱混响条件下的分离信号输出信噪比平均达到了16dB。

Underdetermined Blind Separation of Delayed Sound Source in the Time-frequency Domain

For the attenuation and time-delay mixture model, a new speech blind separation algorithm based on sound source attenuation and time-delay estimation via nonlinear time-frequency masking was proposed. Omitting the matching procedure between attenuation and time-delay, the number of sources and sound source azimuth were estimated directly using the potential function plane clustering by virtue of the approximate sparse criterion of speech signals. The new potential function proposed was more adaptable. In the procedure of separation, the object speech was separated via nonlinear time-frequency masking which was based on the reconstruction of the transfer function by making use of the sound source azimuth..Sound sources located in different position were successful separated .Simulation results showed that our proposed speech separation algorithm can lock interested object speech from random direction and extract object speech effectively, the output signal-noise-ratio (SNR-out) was obtained 16dB averagely in our experiment condition .