Speech and Music Classification Using Hybrid Form of Spectrogram and Fourier Transformation

This paper presents the technique for feature extraction to classify speech and music audio data. The combination of image processing and signal processing is used to classify audio data. There are three main steps. First, the audio data is segments and transformed to spectrogram image and then apply image processing methods to find the salient characteristics on the spectrogram image. The next step transforms the salient spectrogram image using 2-dimensional Fourier Transform and then calculates the energy of signal at the specific frequencies to form the feature vector. Next, in classification process, Support Vector Machine is used as bi-classification tool. The method is tested on an audio database containing 510 instances with 1.5 seconds length of each. The experimental results show that the acceptable classification accuracy of our proposed technique is achieved.