Instrument Identification Phase Space Reconstruction

Along with the development of network technology, multimedia terminaland wireless transmission technology, the audio devices become ubiquitous inour life, and the demand of general public for music increases in a geometricalratio. Many applications based on the music signals set a higher request to theaccuracy of the analysis results, such as music score translation, automatictagging and music retrieval, etc. As is known to all, music signals needinstruments to play. Instrument recognition plays an important role in the filedof music signal processing. More and more experts and scholars focus theirattention on the research of the instrument recognition. It will become a researchhotspot in the field of audio processing.With the objective of improving accuracy and robustness of instrumentrecognition, this article studies the phase space reconstruction model based onprincipal component analysis, the theory of flexible neural trees andidentification model construction, etc. When analysis in feature extraction of theinstrument signals, in view of the traditional feature extraction have focused onthe problem of the time domain space, in this paper we propose a phase spacereconstruction model based on principal component analysis. Themultidimensional phase-space is reconstructed, it kept the equivalence with themotive force system in the topological sense, it contains the dynamics ofinformation system is also more obvious. Then uses principal componentanalysis to reduce the dimensions of rebuild signals, and through theimprovement of the phase space reconstruction model to further reduce thedimensions of the feature set, and reduce the computational complexity in thefuture. In instrument recognition model, we introduce probability densityfunction method into feature extraction, for different instruments, it has differentphase space reconstruction graph, so that, the differences between them can bedescribed. To the variability of instrument signal, this paper we adopted theflexible neural tree as a classifier. Using the flexible tree structure, we can automatic selection of input variables and reduce the dimension of input space.It can improve the adaptive ability of various target classification problem. Theexperimental simulation results show that the proposed method has a goodrecognition result.