Performance comparison for feature selection in Musical Information Retrieval

Music Information Retrieval (MIR) is an interesting area of investigation. The MIR research aims to develop new techniques for processing musical information and searching music databases by content. Therefore, robust retrieval and matching techniques are required. This paper devises a more practical and efficient approach to MIR by investigating a variety of statistical and signal processing-based features, such as Fast Fourier Transform (FFT), Linear Predicative Coding coefficients LPC, Pitch and the wavelet analysis. The features were tested by using different measures of melodic similarity to achieve a better search in musical databases. The paper uses the recall measure to evaluate the retrieval results. It indicated a poor retrieval quality with statistical features and LPC parameters and it is improved when we used the FFT coefficients, and finally using wavelet coefficients caused a significant improvement in its value.