Multiple Pitch Estimation For Music Signal

Multiple Pitch Estimation (MPE) of polyphonic music is one of the most important and difficult issues in the area of the Music Information Retrieval (MIR). The main job of MPE is to estimate both the accurate value and the number of the fundamental frequencies. The main purpose of this thesis is to balance the contradiction between the lower computation cost and the higher accuracy of MPE. The main contents and innovations are shown as follows.(1) In order to show the current situation of MPE, this thesis comprehensively summarizes the MPE algorithms in time domain, frequency domain and time-frequency domain and finds the performance of iterative spectral subtraction algorithm and joint estimation algorithm are better than others. This thesis concludes the physical reasons and the solutions of the three key issues of the harmonic missing, the harmonic overlap and the octave error. In addition, two evaluation criterions are selected to evaluate the provided algorithms and the referenced algorithms.(2) The Harmonic Product Spectrum (HPS) is used to estimate the fundamental frequency candidate set. The Partial Rearrangement (PR) and Octave Confirmation (OC) algorithms are developed to solve the problems of harmonic overlaps and octave errors. The PR algorithm can separate the overlapped harmonics caused by concurrent notes. The OC algorithm can correct the sub/supper pitch error that is in the range of six octaves indicated by vector [1/4,1/3,1/2,2,3,4]. Based on single frame, this thesis proposes a method which consists of pre-processing and post-processing, in addition to the above three discussed procedures. The pre-processing, which uses the HPS to estimate FO candidate set, largely reduces the computational complexity. The PR and OC models greatly improve the estimation accuracy. Therefore, the proposed method realizes the balance between the computation cost and estimation accuracy.(3) Chord recognizer, which uses the interval relationship between notes of in-situ chord and translocation chord is utilized to refine the fundamental frequency candidate set. This thesis also proposes a MPE method based on multi-frame and chord recognizer and Hidden Markov Models (HMM). The88notes which range from27.5Hz to4186.0Hz are all modeled into HMM of two states. Though the HMM and Viterbi algorithm raise the computational cost, chord recognizer and HMM, which emphasize the time continuity of notes, improve the estimation accuracy, apparently.Experiment results show that, the proposed methods based on single frame and multi-frames both have higher accuracy and lower computation cost on both the random note combinations and the music clips. Based on the selected two evaluation criterions, the proposed two methods both exceed the iterative spectral subtraction algorithm and joint estimation algorithm.