基于Mogrifier技术的音乐源分离模型的实现

Implementation of Music Source Separation Model Based on Mogrifier Technology

近年来,互联网和大数据的发展促使语音分离技术得到提升,但是单声道音乐源分离的发展比较缓慢。现有的分离模型存在很多问题,泛化能力和语言建模能力有待加强。为了提高模型的泛化能力,对UMX进行改进后构建了一种通过交互计算在深层次上可以提取音频特征的神经网络结构(M-UMX)。M-UMX可以更好地建模输入和当前状态的关系,能够捕捉到时间序列更长的信息从而优化梯度流在整个神经网络中的传递。实验表明,在MUSDN187S数据集上,相较于UMX,M-UMX建模的泛化能力得到了进一步提升,有效提高了音乐源分离性能。

In recent years, the development of the Internet and big data has promoted the voice separation technology, but the development of mono music source separation is relatively slow. The existing separation models have many problems, and the generalization ability and language modeling ability need to be strengthened. In order to improve the generalization ability of the model, a neural network structure(m-umx) which can extract audio features at a deep level through interactive computing is constructed after improving UMX. M-umx can better model the relationship between input and current state. It can capture longer information of time series, so as to optimize the transmission of gradient flow in the whole neural network. Experiments show that the generalization ability of m-umx modeling is further improved compared with UMX on musdn187 s data set. The performance of music source separation is effectively improved.