Research On Western Instrument Timbre Recognition Based On Monophonic Instrument Sample

As an important branch in the field of music information retrieval,the key to musical instrument recognition is the extraction of timbre information.According to the sounding mechanism of instrument,the timbre of musical instrument is expressed as the time evolution of different frequency components of the musical tone.Traditionally,the carriers of the timbre information are artificially designed features,and therefore the stable timbre component of musical instrument is represented by the evolution of time sequence of timbre feature.The research work of this thesis focuses on the timbral recognition of Monophonic music of Western instruments.Firstly,in this thesis,after the feature selection and dimension reduction,different combinations of timbre features are put into the shallow classifiers to recognize different musical instrument.Among them,the Gaussian Mixture Model and the Universal Background Model express the musical instrument timbre as probability distributions,while the Hidden Markov Model express musical instrument timbres as the probability distributions and transition probabilities of hidden states.Regardless of which one of the three classifiers,the feature selection method based on the Information Gain(IG)principle exhibits superior performance.The Universal Background Model outperforms the other two classifiers and achieves an overall accuracy of 92.3%.In addition,the time integration of the timbre feature series and the Support Vector Machine are used to realize the instrument timbre recognition.For different timbre features,the combination with its time integration,CMAR,shows advantages over simply MAR.At the same time,time integration of MFCC shows significant advantages with the overall accuracy rate of 92.3%.This phenomenon stems from the defects of artificially designed timbre features.In the both cases with and without time integration,the recognition performance of wind instruments is not as good as that of string instruments.Secondly,in this thesis,the timbre feature time integration is also put into the deep neural network classifiers to compared the performance with SVM.The deep classifiers greatly improve the weak recognition of wind instruments.At the same time,the deep classifer also improves the overall performance of the instrument recognition and suppresses the confusion between the instruments or the instrument families.Similar as the recognition using SVM,CMAR shows advantages over the pure MAR feature.MFCC again shows significant advantages.Regardless of the timbre features,the deep model clustering effect given by t-SNE shows that:(1)As the number of layer of deep models increases,the two-dimensional vectors corresponding to different instruments present both the intra-class aggregation and the inter-class separation;(2)After combining the time integration,the aggregation effect of each layer is clearer.Meanwhile,the clustering effect is consistent with the analysis of the recognition measurement.The Convolutional Neural Network(CNN)achieves the highest overall accuracy of 99.57%;while the Deep Neural Network(DNN)reaches 99.02%.Though the deep classifiers efficiently refine the significance of the chosen timbre features through the non-linear transformation layer by layer,they do not completely overcome the insufficiency of the selected timbre features.Finally,the extra timbre feature is extracted directly from the music singal waveform by Deep Convolutional Auto Encode(DCAE)and evaluated by the instrument recognition measurement of CNN and DNN models and t-SNE clustering.The evaluation results show that the unsupervised training helps to optimize the supervised training,which is embodied by the improvement of wind instruments recognition.