Researchs On Music Chord Recognitions

With the improvement of available internet bandwidth,and the continuous development of multimedia information compression technology,the storage and release on internet music are becoming more common.To meet the needs of the customer to retrieve the music anytime and anywhere,content-based music retrieval emerges as the times require.As a typical mid-level feature,chord contains rich information which can represent musical attributes,and has a very important role in analyzing the structure and melody of music.In this thesis,the technology of chord recognition on music is explored,the robust representation of music chords feature and two chord recognition algorithms are proposed.In this thesis,the sequential sparse representation classification and sequential support vector machine method on chord recognition are proposed combined with music theory,signal processing,pattern recognition and other relevant knowledges.Based on signal processing,the main research includes features extraction and chord estimatation.The main contributions in this dissertation are as followings:(1)A robust log pitch class profile(RLPCP)feature is introduced.One of the key to chord recognition is the feature.LPCP is based on the beat pitch class profile and can better express the content of music audio?And it can improve the chord recognition rates.In order to reduce the influence of voice,singing voice separation is accomplished before the calculation of PCP.This enables the accompaniment containing cleaner features,and thus the music audio files are robust for chord recognition.(2)A method named sequential sparse representation classification on chord recognition is proposed.Chords of input audio clips are estimated with the train samples.Sequential sparse representation model is combined with hidden markov chain model and overcoming disadvantages of training model parameters.The recognition model is evaluated on the MIREX'09 dataset for Audio Chord Estimation.The recognition rates of the methods recommended in this thesis are higher than that of the the Music Information Retrieval Evaluation eXchange(MIREX)competition from 2013 to 2015 in audio chord estimation using the MIREX'09 dataset for the MIREX Audio Chord Estimation task.(3)Sequential support vector machine on music chord recognition is suggested.In order to overcome time-consuming of sparse representation classification,the support vector machine is used for chord recognition.Once the model parameters are trainied in advance,time for chords estimated is shorter.At the same time,sequential support vector machine model takes full advantage of the characteristics of music chords changing in time domain.