Presenting and classification based on three basic speech properties, using Haar wavelet analyzing

Due to the importance of speech signal in communications, feature extracting and classification of speech based on important attributes of this signal have became a priority. In this paper, a set of extracted speech features is discussed. The language of speech dataset was Farsi with emotional states such as happiness, sadness, interrogative and normal. In this way, three features (i.e. zero crossing rate (ZCR), standard deviation (SD), and average magnitude) are extracted, using Haar wavelet. For this purpose, first the speech signal is divided into five sub-layers, using Haar wavelet and the mentioned features are extracted for each of these sub-bands. Then, the extracted data is classified using support vector machine (SVM) algorithm. In this way, radial basis function (RBF) kernel function is used because of nonlinear relations in data. Also, two methods have been used in classification: one versus of the rest and pair-wise (couple). Empirical results show that the correct classification rate of test data is about 89% when using pair-wise method. For one versus of the rest method, this rate is decreased to 67%.