LAUSR

Of the many audio features available, this paper focuses on the comparison of two most popular features, i.e. line spectral frequencies (LSF) and Mel-Frequency Cepstral Coefficients. We trained a feedforward neural network with various hidden layers and number of hidden nodes to identify five different languages, i.e. Arabic, Chinese, English, Korean, and Malay. LSF, MFCC, and combination of both features were extracted as the feature vectors. Systematic experiments have been conducted to find the optimum parameters, i.e. sampling frequency, frame size, model order, and structure of neural network. The recognition rate per frame was converted to recognition rate per audio file using majority voting. On average, the recognition rate for LSF, MFCC, and combination of both features are 96%, 92%, and 96%, respectively. Therefore, LSF is the most suitable features to be utilized for language identification using feedforward neural network classifier.