Obtaining the feature vectors of Turkish musical instruments and classification using artificial neural networks

The main aim of this thesis is to analyze the timbre characteristics of various Turkish Musical instruments in digital environment in addition to discriminating and classifying these instruments. It is aimed that this discrimination process, which is sometimes difficult even for the human ear, is performed with high performance. This classification is thought to be benefical in several respects: - Since the classification in this thesis is based on the relation of each instrument with other ones, it is possible to deduce the structural characteristics of an instrument from its sound recording. - It is possible to examine recordings with single instrument (taqsim etc.) more clearly in terms of maqam theory in the light of numerical datas. - It is hoped that for recordings with more than one instrument, it will be a starting point and source of inspiration in order to determine which instruments take place in the recording and to determine the notes played by each instrument. In the first part of the thesis, information about the sound systems which is used in Turkish and Western Classical Music was given and the differences between the notes were tried to be explained. In the second part, the characteristics of the recording medium are explained and the short history of the recorded Turkish musical instruments was introduced. The recordings were taken in the recording studio of the Istanbul Technical University Conservatory Music Technologies Department after the appropriate microphone technique was specified for each instrument. Under the same ambient conditions, a basic two-step method was followed in order to be able to categorize each musical instrument after taking the sound record of all note values in a chromatic octave sound field. First of all, the sound recordings are transformed into datas which reflect the acoustic properties of the instruments using certain mathematical methods and then they are classified by various artificial neural networks including learning algorithms. A general information is provided about mathematical methods that are used for the third and the forth stages of the thesis and the application results are shown in the fifth and the sixth stages of the thesis. Statistical parameters such as Cepstrum, Fast Fourier Transform (FFT), Autoregressive Model (AR), time domain statistical parameters (TDSP) and Discrete Wavelet Transform (DWT) have been employed as the method of obtaining the attribute vectors. The datas obtained has been used as an input vector on the Multilayer Feedforward Artificial Neural Network. According to the methods of backpropagation (BP), Levenberg-Marquardt (L-M) or Gradient Decent (GD) approaches have been employed in Artificial Neural Networks (ANN) according to the style of the feature vector. When the size of the input vector decreases, the Levenberg-Marquardt method is preferred because the working speed increases, whereas the GD method is used for the ANN with the large input vector. In other words, GD method has been used for the datas obtained with the cepstrum and FFT while both GD and L-M methods were used for the datas obtained with the time-domain statistical parameters, AR and ADD models. The success rates of the obtained results were compared. In general, it is recommended that multilayer feed-forward networks use more than twice as many hidden layers as the number of input. However, as the number of inputs in the analysis made in this thesis is considerably high, and a considerable capacity of processor is required in order to operate such a program, it was decided to use 40 neurons in the hidden layer of the network. In the output layer, there are two neurons, one of them classifying the instrument and the other trying to identify the value of the note. In this study, sound recordings were made at four different frequency values, namely rast, zirgüle, neva and gerdaniye for 6 Turkish Musical instruments (ud, kanun, classical kemençe with three strings, bağlama, ney and tanbur). And then feature vectors were obtained implementing each of the above mentioned methods, and they were tried to be classified using multilayer feedforward artificial neural networks. As a result of the analysis, the most successful classification is obtained by documenting the attribute vectors with FFT and giving them as an input to the backpropagation ANN with the help of Gradient Decent algorithm. Applying the data obtained from the Capstrum analysis to the artificial neural networks with the same learning method as the input vector is more successful than the discrete wavelet transform in the prediction of the notes but it is less successful the discrete wavelet transform in the instrument detection. It can be said that success in the second rank after HFD is achieved by these two approaches. The Levenberg Marquardt method yielded better results for the autoregressive (AR) model, which is one of the five methods to be tested. The least successful method has been the processing of feature vectors containing statistical parameters of the signal such as mean value, standard deviation, skewness and kurtosis.For this input vector, it has been observed that Levenberg-Marquardt methods were more successful in ANN. In conclusion, the comparison charts generated in the light of this obtained information are given in the final section of the thesis.