Mel Pitch Class Profiles Feature In Music Chord Recognition Algorithm Application Research

Music is an important kind of audio data.The traditional music retrieval system is based on the key words, which limits the practical applications.With the rapid development of the internet and the popularization of digital devices,rapid and effective searching for desired audio data becomes an important and challenging research topic. In this dissertation, author researches the key technologies of content-based music retrieval in-depth,and proposes recognition methods for piano single-note,representation of music chords feature ,and recognition algorithms.The main contributions of this dissertation are summarized as follows:(1) Based on the study of the development of computer music and basic music theory,author define the classification of the music features,the music features are divided into: basic features,part features and whole features which are the foundation for the next music recognition.(2)According to music and acoustic theory, combining with characteristics of speech signals, author select Mel frequency cepstrum coefficient as the characteristic of monophonic music signal .We also discusses the selection of dimensions of feature vector. We recognize the eighty eight tones of piano with the method of RBF Neural Network and Supported Vector Machine. The final recognition rate is 100%, which show that the methods we selected is effective for monophonic music recognition.(3)The paper studies the Pitch Class Profile features and the calculation method which is often the choice of the feature in automatic chord recognition, in western mucic, calculation formula is list out. Aim at fuzzy information in low-frequency of PCP features and combine with human auditory characteristics, we proposed a new chords feature:MPCP(Mel PCP).We construct the music chord recognition system based on HMM, define 36-state for the HMM, each state represents a single chord. The observation distribution is modeled by a single multivariate Gaussian in 12 dimensions defined by its mean vector and covariance matrix. We get the supervised HMM by virtue of the label files made by Chris Harte. In the course of recognition, viterbi algorithm is applied to the model to find the optimal path,i.e. chord sequence, in a maximum likelihood sense given an input signal. We achive recognition system applying with the two features apart, and analyse the results of recognition.(4) In order to ensure correctness of math derivation, the recognition algorithm of HMM has strict independence assumption, relativity of observation value being limited.We proposed the music chords recognition system based on the Conditional Random Fields(CRFs). The experiment result by HMM and CRFs indicate that the system based on HMM is less than CRFs in mathematics and trainning time, but the overall recognition rate is far behind the CRFs.(5)Author study the framework and principles of Cultural Algorithm(CA).aim at problems of parameter estimate and the long training time of CRFs, author proposes a new model: CA—CRFs that embed CA,which be applied with chords recognition.The result of new model not only improves the recognition rate,but also shortens the time.(6) For effect of individual number in the evolution process in CA, author propose to optimize individual number using fuzzy controller,and combind the improved CA with CRFs, we establish a CA-CRFs chords recognition system based on the fuzzy CA.(7)For fuzzy rules faultiness and control process being nonlinear,we prorose to control accept function of CA by adaptive fuzzy controller, optimize CA. Finaly we establish a CA-CRFs chords recognition system based on the adaptive fuzzy CA,which enhances memory of CRFs and improves recognition rate.