Music Note Onset Based On Phase Features

Along with the development of multimedia technologies, music data volume was getting more and more huge. It is required to classify and retrieve these resources reasonably, but the accuracy of note detection in music is the key of entire system. Nowadays the technology of note onset detection is at the start stage, mostly profited from the method of speech signal processing, for example, methods of short-time average energy, zero crossing rate examination and so on, which were based on the research of frequency feature. This thesis explores the extraction of notes'phase features, and proposes a new detective method. It combines with the theory of all-phase preprocessing and the feature of notes changed by accident phase deviations.This thesis firstly has made various understanding to the note onset detection's research background, and summarizes the current detection in three important steps: preprocessing, reduction and peak-picking. Then explains the each step in detail, that including kinds of current correlation theories and approaches. In the end, the thesis has made the simple analysis and contrast on them.Secondly, the thesis considers that apFFT not only could extract signals'phase features accurately, but also could suppress the spectrum revelation effectively. So it chooses all-phase preprocessing to filter music signals. After that, the thesis uses the method of phase deviation to detective the note onsets of music. Then the peaks of onsets are indicated by Median-filter processing.At last, the thesis lists the results of the experiment on massive music segments, separately using the approaches of short-time average energy, zero crossing rate examination, high frequency content (HFC) function and phase deviations function. Viewing the obtained preliminary figures, the former two methods are simple and inaccurate; HFC is successful at emphasizing the percussiveness of signal, but less robust at detecting the onsets of low-pitched and nonpercussive events; Phase-based methods are successful at detecting low and high frequency tonal changes, but suffers from variations introduced by the phases of noisy low-energy components.