Gaussian mixture model classification of odontocetes in the Southern California Bight and the Gulf of California
Acoustic recordings were made along the transect and track lines in (a) the Southern California Bight and (b) the Gulf of California. The southern California data were collected on two series of cruises, one of which concentrated on the area contained in the square region.
Flow diagram for feature extraction. Overlapping frames of are taken from the signal and transformed to the cepstral domain.
Spectrograms and cepstral coefficients for (a) and (b) whistles, (c) and (d) clicks, and (e) and (f) burst pulses. The left column shows a spectrogram, and the right column shows the corresponding cepstrogram of the same signal after the application of a 64 point DFT based filter bank. The whistles are produced by bottlenose dolphins, the clicks and burst pulses by Pacific white-sided dolphins.
Illustration of (a) spectrogram and (b) cepstrogram for multiple overlapping calls from a group of common dolphins. The classifier seeks patterns associated with multiple calls and does not attempt to separate out individual calls. The effects of a linearly spaced 64 band filterbank are shown in spectrogram (c). The cepstral series can be truncated while still retaining much of the original information as seen in the reconstruction of (c) using the first 32 cepstral coefficients of (b).
Classification of sequences of feature vectors by Gaussian mixture models.
Classification accuracy using Gaussian mixture models with differing numbers of mixtures. The circles represent the percentage of segments correctly classified and their 95% confidence intervals. The triangles show the means of the classification rates across species. Considering only overall accuracy would bias the classifier toward the species with the greatest number of test utterances.
Accuracy by species as the number of mixtures per GMM varies.
Classification accuracy when varying the length of test segments classified by 256 mixture GMMs. The circles represent the percentage of segments correctly classified and their 95% confidence intervals. The triangles show the means of the classification rates across species.
Accuracy by species as the length of the test segment classified by a 256 mixture GMM is varied.
Number of seconds of usable call data obtained for each dolphin species by date.
Selection of train/test data. The listed sessions from Table I are used for training with the remaining sessions used as test. Partitions 2–4 were chosen such that the training data for a single species is replaced by one of the other sessions.
Performance of 256 mixture GMMs with test segments and varying amounts of training data per species.
Results of 256 mixture GMM tests with test segments on the partitions resulting from Table II . Column CI is the 95% confidence interval on the overall accuracy, and species mean represents the mean of the individual species’ accuracies.
Article metrics loading...
Full text loading...