With the rapid growth of digital music today, due to the complexity of the music itself, the ambiguity of the definition of music category, and the limited understanding of the characteristics of human auditory perception, the research on topics related to automatic segmentation of music is still in its infancy, while automatic music is still in its infancy. Segmentation is a prerequisite for fast and effective retrieval of music resources, and its potential application needs are huge. Therefore, topics related to automatic music segmentation have important research value. This paper studies an improved algorithm based on negative entropy maximization for well-posed speech and music separation. Aiming at the problem that the separation performance of the negative entropy maximization method depends on the selection of the initial matrix, the Newton downhill method is used instead of the Newton iteration method as the optimization algorithm to find the optimal matrix. By changing the descending factor, the objective function shows a downward trend, and the dependence of the algorithm on the initial value is reduced. The simulation experimental results show that the algorithm can separate the source signal well under different initial values. The average iteration time of the improved algorithm is reduced by 26.2%, the number of iterations is reduced by 69.4%, and the iteration time and the number of iterations are both small. Fluctuations within the range better solve the problem of sensitivity to the initial value. Experiments have proved that the new objective function can significantly improve the separation performance of neural networks. Compared with the existing music separation methods, the method in this paper shows excellent performance in both accompaniment and singing in separated music.