A Research On Music Source Separation

Music Source Separation refers to the problem of extracting each single instrument sound or some specific instrument sounds from the mixture. It arouses more and more interests in recent years since Music Source Separation technology plays such a significant role in music instrument recognition, music melody extraction, content related music indexing, automatic music separation, automatic music transcription, etc.This paper briefly introduces Music Source Separation relative foundation theories, sinusoidal analysis & synthesis technology and separation performance assessment methods at the beginning. In this paper, the music source separation techniques are classified into two categories: streaming algorithms and de-mixing algorithms, which are followed by an introduction of the main ideas of each category respectively. For streaming algorithms, sinusoidal model based algorithm, CASA (Computer Auditory Sense Analysis), and KNN (K Nearest Neighbor) are discussed in detail whereas for the de-mixing algorithms, this paper focuses on spectral filtering based music source separation and NMF based music source separation.Sinusoidal model based algorithm discussed here is a traditional approach. Peaks matching module combines the integer multiple harmonic tracks with an improved algorithm considering both the pitch and the amplitude. The experiment results demonstrate that the improved sinusoidal algorithm has better performance than the traditional sinusoidal algorithm. CASA based music source separation algorithm utilizes Gammatone filter and masking effect to separate music signal. The KNN based music source separation algorithm utilizes prior information of music source which can also achieve music source separation.Spectral filtering based separation algorithm discussed in this paper designs filter based on multipitch detection, aiming to separate music signal from two aspects, namely, decimation-in-frequency and time-domain filtering. The typical NMF (Non-Negative Matrix Factorization) algorithm can't ensure the independency of basis spectral which arouses the distortion of separated music source. To solve this problem to some degree, this paper purposes an improved NMF algorithm. This paper simulates the five separation algorithms discussed above and assesses their performances from subjective evaluation, objective evaluation and overall evaluation respectively. This paper utilizes MOS (Mean Opinion Score) as the subjective evaluation. The three objective evaluation indicators are SNR (Signal-to-Noise Ratio), CC (Correlation Coefficient), and Kurtosis.