Abstract: To improve separation performance of musical instrument separation algorithm in a noisy multi-track polyphonic domain，4th cummulants are introduced which measured energy nonstationarity of instrumental time series.With time-delayed autocorrelation，a combined blind musical instrument separation algorithm is presented.Firstly，using 4th order cumulants，compute linear transformation through which globlal mixing matrix is orthogonalized with additive noise.Then the mixing matrix and source instrument signal can be estimated from the preprocessed multi-track musical data via the 2nd and 4th order combined time delay statistics.Experiment demonstrates that the combined algorithm can separate source instruments in polyphonic signal more precisely than which only using 2nd order statistics，especially in lower SNR environment.

Key words: time delay 4th order cumulants, blind musical instrument separation, time delay autocorrelation, robust orthogonalizing

摘要： 为改善有噪条件下多轨复调音乐乐器分离算法性能，引入了度量乐器时间序列能量非平稳性的四阶累积量，提出了一种二、四阶组合时延统计量复调音乐乐器盲分离算法。首先根据四阶累积量函数，得到使全局混合矩阵与加性噪声正交的线性变换。对经线性变换预处理后的多轨音乐数据利用二、四阶组合时延统计量，估计出混合矩阵和源乐器信号。仿真实验证明，这种组合时延统计量乐器分离法较二阶统计量算法能够更精确地分离出复调音乐中的源乐器，尤其适于低信噪比情况。

关键词: 时延四阶累积量, 乐器盲分离, 时延自相关函数, 稳健正交化