WARSHIP SOUND SIGNATURE RECOGNITION USING MEL FREQUENCY CEPSTRAL COEFFICIENTS

Keywords: Sound source recognition, signal processing, support vector machine, cepstral coefficients, warship acoustic, feature extraction, MFCC.

Literature survey about sound source recognition/classification systems shows that there are studies about music, musical instruments, human and vehicle sounds but none about ship sounds. In this thesis, a system which recognizes the type of a ship by evaluating the sound of working machinery on it (engine, generator, propeller etc.) was developed.

First the interferences are cleaned by Low Pass Filter (LPF). Then the sounds are divided into frames using Hamming window. For each frame 13 and 9 coefficients are extracted by Mel Frequency Cepstral Coefficients (MFCC) and Delta Mel Frequency Cepstral Coefficients (DeltaMFCC); 6 coefficients (minimum, maximum, mean, median, standart deviation and range values of pitchs) are extracted by Pitch Detection. Except the first coefficient which represents the total energy density and because of this it is so far from the others, remaining (12/8) coefficients of MFCC/Delta MFCC are used. For 10 second ship sounds, 992-1009 frames are formed and these coefficients are reduced to [12*64] matrix for recognizing by vector quantization method. The system is trained using k Nearest Neighbour (k-NN) and Support Vector Machine (SVM) methods with the training set which consist of 90 different sounds recorded at different speeds of 12 different ships. Because of the ship recognition process is a multi-class problem; SVM’s one-against-one (pairwise) approach is used for recognition.

The system was tested with 110 different ship sounds and the true recognition rate was 82% with MFCC (12 coefficients) and SVM.