(Color online) Example of tonal sound detection process: (a) Original spectrogram. (b) Spectrogram after pre-whitening and smoothing. (c) Tonal spectral peaks after extraction. Empty circles represent the tonal spectral peaks excluded by isolated noise filter. Filled circles represent the adopted frequencies. Three examples of occupancy rate (O) were calculated within the examined range (empty square).
Example of tonal spectral peak extraction. (a) Power spectrum after pre-whitening and smoothing from the time slice at 0.14 s in Fig. 1 . Filled and open circles represent the amplitude deviation (S) ≥ 1 and < 1 separately. (b) Spectra of second derivatives. Filled and open circles represent the second derivative (D) ≥ 2 and < 2 separately. (c) Power spectrum after filtering by S th and D th, filled and open circles both represent the data fulfill with the previous two conditions. But only one tonal spectral peak (filled circle) is extracted in this example because the instantaneous frequency bandwidth is 5Δ f .
(Color online) Example of detection result. (a) Original spectrogram. (b) Scatter plot of adopted frequencies that belong to fundamental frequency (filled circle) and harmonic (open diamond). (c) Percentage occurrence of harmonics and overlapping tonal sounds calculated within 0.1 s. (d) Short time Shannon entropy calculated within 0.1 s based on the adopted frequencies belong to fundamental frequency.
ROC curves for the recordings of humpback dolphins and the MobySound archive yielded by varying the threshold of occupancy rate (O).
Distribution of SNRs for all 1 s segments.
Audio files corresponding to the summary results of Table II and III .
Summary on the detection results of the MobySound Archive recordings. The 25th, 50th, 75th percentiles of adopted frequency (AF) were calculated based on the data after excluding harmonics. The 50th percentile of entropy was calculated based on the short time entropy in 0.1 s with the percentage of overlapping tonal sounds < 20%.
Summary on the detection results of the humpback dolphin recordings. The 25th, 50th, 75th percentiles of adopted frequency (AF) were calculated based on the data after excluding harmonics. The 50th percentile of entropy was calculated based on the short time entropy in 0.1 s with the percentage of overlapping tonal sounds <20%.
Summary on the deviations of frequency distribution extracted from our algorithm and ground truth data. Bold font showed the mean deviation is significantly different from 0. The mean and standard error (S.E.) of the deviations on the five percentiles of adopted frequency are given in Hz.
Article metrics loading...
Full text loading...