Multipitch Estimation Of Polyphonic Music Based On Sparse Decomposition

Music which civilizes and inspires men everywhere is one of the most significant parts in our daily life. And the analysis of music signal plays an active role in the course of music research. Pitch as a major part of music information can be detected and implemented in automatic music transcription, music separation and music retrieval etc.Many signal processing technical which base on time domain or frequency domain point have been used in estimating fundamental frequency of mixed music signal. But all of current methods have more or less shortcoming because of the inharmonic of high partials, fundamental frequency component may be disappeared, the interval of pitch is too small in bass line, the overlapping of spectrum in different notes.Although, A lot of methods was posed from different view for solving the difficulties in extracting fundamental frequency. For example, the filter which corresponds to human's auditory sense can be used to suppress inharmonic components. And bass notes could be detected by period or using a higher frequency resolution than melody notes.The main content of this thesis was summaried as follows:1,Sparse decomposition is a newly emerging technology with great advantage in signal processing. It is leaded to the fields of Multipitch estimation for music signal processing in this thesis. According to sparse property of pitch, the match pursuit (MP) is applied for extracting the spectrum of notes based on pitch definition and music model.2,A dictionary which simulates human's auditory sense can be built by bring in ear frequency response and multi-resolution technology. Obviously, the MP coefficients response to auditory sense and a unique expression what means auditory frequency of notes has been gotten.3,In the light of music synthesis theory of vibration chord, the energy of fundamental frequency component and its harmonic partials is exponent attenuation and the timbre known as spectrum of note can be extracted from MP coefficients. So a notes dictionary could be built by using sinusoidal model which is accepted widely in academic.4,The mixed signal is first simplified with MP decomposition on auditory band, then reconstructed and decomposed to notes dictionary by base pursuit (BP) algorithm. BP coefficients describe the composition of notes so all of pitches are extracted. The simulate test shows it works pretty well.