Research Of Music Chord Recognition Based On Robust PCP And Metric Learning-Support Vector Machine

As to the massive music resource on internet,the traditional text tag based music information retrieval method can't satisfy the need of deep retrieval of users and the need of application in music content.Content-based music retrieval application is widely used.Chord is an important concept of mid-level music features.Chord contains important music content information,so the automatic transcription of the chord information of music becomes an important topic in the field of intelligent information processing and recommendation system.Automatic chord transcription plays an important role in automatic accompaniment,Internet song songs genre and emotion detection,classification and other fields.Combined with related knowledge of music theory,intelligent information processing and machine learning knowledge,the thesis introduces an algorithm of chord recognition based on metric learning based Support Vector Machine and robust Pitch Class Profile feature.The thesis includes the following aspects:At first,the thesis proposes a new feature extraction algorithm which combines the nuclear-norm convex optimization with Pitch Class Profile(PCP)since the recognition rate of the chord is low when large and sparse noise exists.Solving nuclear-norm minimization problem will remove sparse noise;at the same time,it will reconstruct the structure of low rank in chord feature,called RPCP.Considering music theory,chord progression has duration time.As a result,the improved RPCP has the property of stability during chord progression,that is to say,we extract the feature based on chord progression other than the method based on frame or beat.Secondly,the thesis applies argument Lagrange Multiplier to solve nuclear norm convex optimization problem.And the thesis introduces a adaptive threshold selecting step to improve time complexity.Thirdly,the thesis applies the metric learning as the training method of support vector machine.Based on train set,metric learning method will obtain metric matrix carrying information of distribution of data set.Then this thesis utilizes this metric matrix to achieve Mahalanobis distance.Then this thesis uses Mahalanobis distance to replace Euclid Distance in kernel function of support vector machine.At last,the thesis describes the chord recognition system based on metric learning based SVM in detail.The total number of chord type is 24 which contains 12 major triad and 12 minor triad.The thesis extracts robust PCP as the new chord recognition feature and realizes the chord transcription and recognition system by metric learning based SVM method.The results show that the ratios of chord recognition has a 5.84% increment after using the robust PCP and metric learning based SVM than using the traditional PCP and SVM.