Research On Multiple Pitch Estimation Based On Piano Note Event Segment

With the development of computer networks and multimedia technologies,music inform-ation retrieval has become a popular research direction in the academic and industrial circles at home and abroad in the field of computer science.The note onset detection and multiple pitch estimation are both important research topics.The foundation of other research directions in content-based music information retrieval is of great significance for further music structure analysis,music retrieval,music style classification,and cover recognition.The research object of this article is mainly piano music in WAV format.Since every key played in a piano performance can be regarded as a note event,this article mainly includes two subtasks:note event detection and multiple pitch estimation of note event segments.The piano music signal is divided into multiple note event segments after the note event detection,and then the number of notes in each note event segment and its fundamental frequency value are extracted through multiple pitch estimation of the note event segment.The main research contents and innovations are as follows:(1)For the method of note onset detection and multiple pitch estimation studied in this paper,most of the methods currently use short-time Fourier transform or constant Q transform as the time-frequency representation in signal analysis.However,the short-time Fourier trans-form has a frequency resolution problem due to the fixed window length.Although constant Q transform has higher frequency resolution at low frequencies,its time resolution is also red-uced.The variable Q transform used in this paper not only has a higher frequency resolution at low frequencies,but also has a better time resolution,so that the music signal has more accurate spectral characteristics after the variable Q transform.(2)Study the algorithm for note onset detection.Analyzing several commonly used methods for note onset detection in the past,it was found that almost none of them have considered those softer note attack,which led to a large number of missed detections.This paper uses power scaling for each frame of spectrum to not only enhance weaker attacks in the music signal but also suppress stronger ones.At the same time,frequency band division weighting is used to enhance the robustness of the detection method;Further through the windowing and difference processing,the correctly detected note attack is increased and the false detection is reduced,thereby improving the overall detection performance,and the average value of the F-measure reaches 96.65%.(3)The multiple pitch estimation algorithm based on non-negative matrix factorization is studied for the note event segment.The multiple pitch estimation algorithm proposed in this paper is based on the decomposition and estimation of the note event segment,taking into account the striking moment,pitch value and ending moment of the note in the piano playing process in advance,and avoiding the false playing error;Taking into account the continuity of music signals and the time-varying nature of notes,the change process of notes with time is divided into transient phase and steady-state phase,and the corresponding note feature dictionary is constructed from the transient phase and steady-state phase of the notes respectively.At the same time,it analyzes the method based on non-negative matrix factorization from two aspects of cost function and sparse constraint,and applies non-negative based on lp,qq norm group sparse constraint to the transient phase and steady-state phase of the measured note event segment and its F-measure reaches 85.22%.