Fig 1.
Spectrogram representations of various bird species showing some of the typical appearances of sounds.
(a) A fox sparrow (Passerella iliaca) song illustrating its syllables, phrases, and elements (S = syllable and E = element). (b)-(e) show representations of lines: (b) tui (Prosthemadera novaeseelandiae); (c) the more-pork sound of ruru (Ninox novaeseelandiae); (d) kakapo (Strigops habroptilus) booming; (e) brewer’s sparrow (Spizella breweri). (f)-(h) demonstrate blocks: (f) (long billed) marsh wren (Cistothorus palustris); (g) female North Island brown kiwi (Apteryx mantelli) call; (h) kakapo chinging. (i)-(j) show stacked harmonics: (i) male North Island brown kiwi whistles; (j) ruru trill. (k) oscillations: North Island saddleback (Philesturnus rufusater).
Fig 2.
(a) A non-stationary signal containing 20 Hz, 40 Hz and 80 Hz frequencies and (b) its power spectrum computed using the Discrete Fourier Transform.
Fig 3.
Examples of bird calls with various degrees of noise, the effect of band-pass filtering and power spectrum of white and pink noise.
The top row of each sound figure displays the oscillogram and the second row the spectrogram. (a) A less noisy example of kakapo chinging with limited noise and (b) a noisy example of kakapo chinging. (c) An original male kiwi whistle and (d) its noise filtered (band-pass) signal. Noise is visible as a grey background in the spectrogram surrounding the sound depiction and most of the high-frequency variation in the oscillogram. Power spectrum of (e) white noise and (f) pink noise.
Fig 4.
Wavelets and their relation to time-frequency resolution and wavelet packet decomposition.
Time-frequency resolution in (a) STFT and (b) wavelets. Examples of mother wavelets: (c) Haar; (d) a subset of Daubechies wavelets; (e) Discrete Meyer wavelet. (f) Scaling and shifting the mother wavelet Ψ1,0(t) gives two new wavelets Ψ2,0(t) and Ψ2,1(t). (g) A level three wavelet packet decomposition tree (A- approximation and D- detail).
Fig 5.
Different mother wavelets produce different results.
Same excerpt of a male kiwi whistle (a) original whistle and (b)—(e) denoised with different mother wavelets.
Fig 6.
An example of kakapo chinging used in the experiment.
Top, middle, and last rows represent oscillogram, spectrogram, and labels indicating the parts of the recording used to calculate the SnNR respectively. (a) Initial signal and (b) the same signal after denoising and band-pass filtering.
Table 1.
List of species, their call types and frequency range.
Table 2.
List of species introduced to the secondary dataset and their song characteristics.
Fig 7.
Denoising different types of noise.
(a) White noise, (b) pink noise, and (c) brown noise.
Fig 8.
Bird call examples of before, after filtering, after denoising using wavelets as described in the text, and after denoising and classical filtering.
(a) A whistle of a male North Island brown kiwi, (b) a call of female North Island brown kiwi, (c) a ruru trill, (d) a ruru more, (e) a kakapo booming, (f) a kakapo chinging, and (g) a less noisy example of male kiwi.
Table 3.
Experimental Results—primary data set.
Fig 9.
Box plot view of the results in (a) Table 3 and (b) Table 4.
Table 4.
Experimental Results for the species introduced to the secondary data set.
Fig 10.
Denoising entire songs and long series of calls.
(a) A North Island kaka song, (b) a marsh wren song, (c) a western meadowlark song, and (d) a series of kakapo chinging.
Table 5.
Comparing the denoising results—series of calls against their segmented calls.
Fig 11.
Box plot view of Table 5.
(a) call series and (b) segmented calls.
Fig 12.
Male kiwi, female kiwi, and more-pork are overlapped in (a) and kakapo chinging overlapped with mottled petrels (Pterodroma inexpectata) in (b).
Fig 13.
A deliberate denoising example.
(a) A North Island robbin song (sampling frequency 44,100 Hz) and (b) its down-sampled song to 22,000 Hz. (c) and (d) are their denoised songs.