Research On Multiple Pitch Detection Of Piano Music

The rapid development of Internet technology allows music to be widely spread. Method for effectively extracting, searching and organizing music information, that is, the study of music information retrieval has been widely concerned by the academic and information community. Multiple pitch detection is one of the hotpots in the field of music information retrieval, and the basis task is to estimate the multiple notes at the same time in polyphonic music, and thus the information of the basis frequency, the begin time and the end time of the notes can be got. At present, the multiple pitch detection method can not meet the actual demand, so it is very important to study futher on the multiple pitch detection method.The thesis takes the piano polyphonic music as the research object. In the framework of multiple pitch estimation method based on the nonnegative matrix factorization, the time-frequency represention of music signal, notes dictionary constructing and spectral decomposition algorithm have been analyzed and improved. And a nonnegative matrix factorization algorithm based on the,l_p,q^q norm term sparse constraint of the multi atomic note dictionary is researched, which effectively improves the accuracy. Finally, a method for estimating the multiple pitch based on the nonnegative matrix factorization is proposed, which is based on the level of notes event but not on the level of signal frame. The main research work and innovation points are as follows:1. The thesis study the multi resolution time-frequency representation constant Q transform（CQT） which is commonly used in music signal analysis, and find that although the CQT has a high frequency resolution at low frequency, it also leads to the decrease of time resolution. We first introduce the variable Q transform as a tool for time-frequency representation of multiple pitch detection, which gets better time resolution in the same frequency resolution and have efficient coefficient calculation camparing with the CQT.2. The thesis study on the spectral decomposition algorithm of single atom and multi atomic note dictionary. By using a new sparse norm constraint term l_p,q^q, the experiment results of the single atom note dictionary shows that the multiple pitch detection with the norm term l_p,q^q is better than the usual norm term 1l. Because the note spectrum changes obviously at different time, it points out that the single atom note dictionary does not take into account the dynamic change of the note spectrum. Then we introduce the construction of multi atomic note dictionary from two aspects: modeling and learning. Finally,l_p,q^q block sparse constrained nonnegative matrix factorization algorithm based on the multi atomic note dictionary is researched. The experiment results show that when the atomic number is 2, the F value obtain 78% for multiple pitch detection of single frame signal in MAPS dataset.3. Because the method of multiple pitch detection methods based on nonnegative matrix decomposition is for single frame signal processing, it does not detect the note onset in advance, but obtains by post processing test results, whilch may get the pseudo onset point and the error of divide a note into multiple notes between the two onset points. A method of multiple pitch detection based on notes event is proposed. Firstly, detecting the note onsets of the music signal, then secondly estimating the fundamental frequency of the musical note event based on NMF, and finally detecting the termination of each note in the note event.