Research On Multiple Fundamental Frequency Estimation Of Automatic Music Transcription System

Automatic Music Transcription (AMT) is defined to convert the pattern ofmusical manifestation from audio waveform to electronic music score throughautomatic musical signal processing. AMT is widely applied to audio content retrieval,audio scenario analysis, music-assistance teaching, music visualization and etc.Multiple Fundamental Frequency Estimation, which is short called Multi-F0estimation,is broadly used for estimation of multiple musical notes, which are pronouncedsimultaneously. But due to the influences of the instruments categories, performingenvironment, players’ habits and indeterminate number of complex tones, theperformance of most of the multi-F0estimations can’t meet the requirements for musiccontent understanding. These influences also have an impact on AMT performancewhich causes the necessity to improve the performance of multi-F0estimation.Based on the analysis and conclusion of existing AMT research achievements,from the perspective of effect and efficiency, the research and analysis of electronicmusic with various styles in this thesis include AMT system model structure, musicsignal pre-process and multi-F0estimation. The main research content and innovativeresults of this dissertation includes:1. It improves the pre-processing method in AMT system which integrates themodel of human audio system into the music signal standardizing process. It usesvolume equalization process instead of normalizing operation of scaled signalamplitude. It has been verified that this method can retain the characteristic of themusic signal effectively and reduce the multi-F0deviation introduced by inconformityof the signal strength.2. It improves the multi-F0estimation by taking the music signal as ordinary,selecting and obtaining the signal characteristics using ordinary signal processingmethod. It chooses an improved Mel Frequency Cepstrum Coefficient (MFCC) as itscharacteristic parameter. Then reduce the signal dimension by principal componentanalysis method. Experimental simulation results indicate that compared with the method taking STFT and MFCC as its characteristics directly, this method can improvethe results for electronic music of which musical notes change equably. This methodhas high efficiency also because time-frequency analysis was used without iterationprocess.3. Based on the tonal and vocal characteristics of musical signal, it improvesmulti-F0estimation methods. Taking harmonic characteristics of a music signal as itsfoundation, this method considers the music signal as specific. It improves the iterativedeleting mechanism, abolish candidate F0selection and reduce the influence offrequency overlap by calculating the harmonic matching ratio from the musical notesthat has the lowest frequency. It has been verified that compared with second iterationdeleting method based on candidate mechanism, this method can improve the AMTeffect for the music that contains no more than five complex tones and is performed byinstruments with resonance characteristic.4. It also proposes a method of multi-F0estimation that based on the combinationof music tonal characteristics and time-frequency characteristics. It integrates musicsignal characteristics into the multi-F0based on time-frequency analysis theory. Firstly,candidate F0is selected utilizing music signal harmonic characteristics. Following that,combination coefficient of the selected results is calculated using time-frequencyanalysis method. The multi-F0estimation results can be obtained. The experimentalresults prove that compared with the two methods mentioned before, it can improve theAMT results for most pop electronic music.5. It proposes another method of multi-F0estimation that based on musiccategories information and probability statistics information. The multi-F0estimationis studied from the perspective of statistics. It obtains the appearance probabilityinformation of every musical note with different music styles. Then it improves theeffects by filtering the previous sorted results based on music categories informationand Bayes statistics information. The experimental results indicate that this method hasa great improvement on AMT results for electronic music with different styles.