Short Time Fourier Transform based music genre classification

Thanks to the development of the computer based decision support systems, many investigations are performed to extract useful information from massive datasets. In particular, music and audio processing are two important topics about data mining and knowledge discovery. In the literature, there are a number of research regarding music signal processing such as music genre classification and music recommendation. To solve these problems, digital signal processing and pattern recognition methods are frequently used. In the proposed study, music genre classification problem has been investigated by using Short Time Fourier Transform (STFT) which is one of the most useful time- frequency analysis methods. In the feature extraction part, spectrogram of the music samples are obtained by using various window types having different window size and overlap ratio. Subsequently, in order to extract the power distribution of the music samples over frequency, time-frequency analysis output is integrated over time. Additionally, the variability of the power for each frequency bin is also calculated and employed as a statistical measure. In order to validate the effects of window type, window size and overlap ratio, Support Vector Machines employing different kernels as a traditional machine learning algorithm and random forest algorithm as an ensemble machine learning approach have been used to classify music samples according their genre.