Numbers of call types are represented according to bins of F-measures for the preliminary experiment. Overall, the F-measures were low, meaning that the classifier (MFCC) was not efficient on the whole dataset.
Mean power spectra are shown for (a) C1, (b) C57, (c) C77, (d) C106, and (e) C113.
Oscillograms (top) and spectrographic representations (bottom) for C1, C57, C77, C106, and C113. Spectrograms were calculated using 512-point Hamming windows, a 22 kHz sampling frequency, and 16-bit amplitude sampling.
Mean percentage correct classification for the five calls selected is represented according to the feature sets. Using just four EDS features provided better results than with 10 or even 20 MFCC values, and the results were even better with more EDS features.
Features composing each EDS feature set. In the feature expressions, “x” represents the input acoustic signal. The mathematical composition operation is implicitly represented by the parentheses. Each label denotes an operator. Most of them are explained in detail in Pachet and Roy (2009). See Appendix for descriptions of operators not included in Pachet and Roy (2009).
Mean percentage F-measures obtained for each feature set. The training-testing process was repeated several times in order to reduce the variability in classifier performance: 50 repetitions (or more) ensure almost constant mean F-measures.
Mean percentage correctly classified calls (±SE) for each feature set and call type.
Confusion matrices providing the distribution of the calls according to their original call type after the classification process with CL2-EDS and CL2-MFCC. More calls were classified correctly with CL2EDS than with CL2MFCC. In both cases, calls of type B were misclassified (in C1, C57, C77, C106 or C113).
Confusion matrices providing the distribution of the calls according to their original call type and mean F-measures obtained for each call type after the classification process with CL1, either followed by CL2-EDS or by CL2-MFCC. More calls were classified correctly with CL2EDS (diagonal of the matrix: 3930 calls) than with CL2MFCC (diagonal of the matrix: 3689 calls). Overall, F-measures were greater with CL2EDS than with CL2MFCC.
Article metrics loading...
Full text loading...