Research On Note Onset Detection Algorithm

As a form of representing the lives and enjoying the art, music has been changing greatly with the rapid development of the information technology, whose storage, production and broadcast have been turning to be digital, centralized and shared. Contend-based music information retrieval is just a new technique and an important research field for these trends. Note onset detection is one of the most important research direction in this field, and is the preliminary work for many other directions. In this dissertation, researches on note onset detection were carried on.First, a new framework of note onset detection inspired by music knowledge base was established after surveying on the existing detection algorithms, and the partial flux feature was proposed based on this framework. The feature connected the partial components of the music with the frequencies of the signal through the twelve-tone equal temperament based on music theory. Because it was inspired by the priori knowledge extracted from the music knowledge base, it was in line with the music psychology, and could improve the performance of the detection algorithm.Second, the partial flux detection algorithm based on the continuous wavelet transform was proposed, for the purpose to overcome the shortcomings of the fixed time-frequency window of the short time Flourier transform in the process of the musical signals processing, and the defects of being difficult to distinguish the musical notes in one octave when using the discrete wavelet transform for the decomposition because of its big span of binary scale. The signals were firstly decomposed by the continuous wavelet transform in the algorithm, and then the partial flux feature was extracted, and the detection function was generated. Finally, the double side exponential smoothing method and the adaptive thresh holding method of moving window normalization were applied to produce the note onsets.Then, the partial flux detection algorithm based on the constant Q transform was presented. The constant Q transform is different from the traditional transform. The idea behind the constant Q transform is that the frequency intervals of the adjacent spectral lines increase exponentially, which is exactly the same as the frequency distribution of the musical partials. The performance was improved by introducing the constant Q transform to the note onset detection algorithm, and the algorithm also had the advantage of low computational cost. Finally, sparse decomposition was employed to the signal transform stage of the note onset detection. The sparse representation under redundant dictionary is able to seize the inherent characteristics of the signal, with excellent adaptability. The matching pursuit algorithm was used as the method of sparse decomposition, and the two novel algorithms based on matching pursuit (MP):the algorithm of MP degree of explanation and the algorithm of MP partial flux, were proposed. Firstly, the musical signals were decomposed through MP, and then the code books were analyzed with the two algorithms. Finally, a modified peak-picking algorithm based on Gaussian kernel smoothing method was applied to generate note onset vectors.The experiments showed that these algorithms were theoretical and practical to some extent. Especially, for the complex mixture music, which is the largest proportion in our real life, the algorithm of MP partial flux had a significant advantage over the other algorithms. The effects of fusion detection algorithm were also explored in this dissertation.All the algorithms mentioned above were programmed using Matlab, and were experimented on the same music datasets, and were evaluated and compared under the international criteria. The experimental results showed that the proposed algorithms were theoretically feasible and practically effective.