Music Source Blind Separation

Music source separation refers to the problem of extracting one or more sounds that human are interested in from mixture of music signals. In a complex sound field environment, the human auditory system can easily separate the interested music source signal from the mixture and keep a high degree of attention. The perfect source separation ability of auditory system promotes the development of study for music source blind separation.In this paper, the problem is divided into two categories:multichannel music source blind separation and single channel music source blind separation. The multichannel mixture model includes determined and underdetermined mixture model. Firstly, by analyzing the statistical distribution characteristics of music signals in time and frequency domains, it is assumed that music signal is super-Gaussian distributed and is similar to Laplace distribution. In the determined mixture model, multichannel music source blind separation is accomplished using the method of independent component analysis (ICA). The separation performance of different algorithms based on different independence indicators such as kurtosis,negative entropy,likelihood are discussed. The result shows that algorithm based on negative entropy achieves the best separation performance for musical signal. Secondly, in the underdetermined mixture model, sparse component analysis (SCA) is used to solve the problem of multichannel music source blind separation under the assumption that music signal is sparse enough in transformed domain. The result shows that the main factor which influences the separation performance is spectra-overlap between sources. Finally, this paper combines nonnegative matrix factorization (NMF) with subspace decomposition theory to solve the problem of single channel music source blind separation. Time-frequency overlapping is considered to be the main factor influencing the separation performance. Aiming at this factor, we add two techniques for spectral and temporal enhancement respectively. The enhanced techniques use the harmonic characteristic of music and similarity of partials time envelopes to inhibit the impact caused by time-frequency overlapping. The result indicates that the enhanced method relieves the interference and improves the acoustical quality of separation result, especially for the pitched instrument.