Research On Automatic Music Transcription Technology

Automatic music transcription(AMT) is the process of converting an acoustic musical signal into some form of musical notation. It is considered to be a key enabling technology in music signal processing,And while it is also a complex and challenging problem. The performance of any transcription system for polyphony music is still significantly below that of a human expert.Automatic music transcription has numerous applications associated with the area, such as automatic music accompaniment, the technology of advanced music editing, Music retrieval, Music Teaching, and so on. Here we divide The AMT problem into two main subtasks, which include: multi-F0 s estimation and multi-F0 s streaming. We also do some research and improve on them. The following is our main research and innovations:(1) Statistical model-based multi-F0 s estimationWe address the multiple F0 estimation problems in a Maximum-Likelihood fashion, where the power spectrum of a time frame is the observation and the F0 s are the parameters to be estimated. When learning the parameters of the models, to take advantage of all areas of the spectrum, we modify the parameters of the model. We add the information of the nonpeak region. The peak region information and the non-peak region information act as a complementary pair. The former helps find F0 s that have harmonics that explain peaks, while the latter helps avoid F0 s that have harmonics in the non-peak regionWhen we estimate the multiple F0 s, to solve the problem of Polyphony estimation, we use the method of hypothetical partial sequences(HPS), calculate the harmonic amplitude and spectrum smoothness, take into account the partials of the other candidates for a given combination. Therefore, we can select the most likely combination according to their energy and smoothness.Finally, we uses a post-processing technique to refine F0 estimates in each frame using neighboring frames, Results show that the improved method shows superior F0 estimation(2) Timbre feature based multi-F0 s streamingWe use the method which based on timbre feature to address the problem of multi-F0 s streaming. We cast the problem as a constrained clustering problem. There we explore different timbre features for music, such as Mel-frequency Campestral Coefficients(MFCC), Harmonic Structure(HS), ordinary cestrum(OC), and Discrete Cestrum(DC). In order to ensure that the same source have similar timbre, and the timbres come from different sources have much more discrimination. We improve the Discrete Cestrum(DC) feature. We called the new timbre feature Uniform Discrete Cestrum(UDC), and it can achieve a better result.When we initial partitions, unlike traditional random initialization,we use the pitch-order initialization. The pitch-order initialization decrease much more the number of iterations for Algorithm while have the similar results. So it can improve the efficiency of the algorithm.