Identification of vocal individuality in male cuckoos using different analytical techniques

Study area and sound recording

Field work was conducted from July 10th to July 17th in 2016, in the Liaohe Delta Nature Reserve (41.033929°N; 121.725244°E), Liaoning Province, northeast China. This region is one of the most important estuarine wetlands in the country and has the largest area of reed-bed habitat along the coastal region of China. Here, the Common Cuckoo is a summer breeding species, and mainly parasitizes the Oriental Reed Warbler (Acrocephalus orientalis) during late May to early August (Li et al. 2016). Using mist nets, we trapped 48 individual cuckoos during the first 3 days of the study. Individuals were immediately banded with a numbered metal band, and marked by waterproof paint applied to the abdomen of the bird to enable observers to distinguish individuals from distance.

We recorded cuckoo vocalizations using a TASCAM HD-P2 portable digital recorder (Tascam Co., Japan) and a Sennheiser MKH416 P48 external directional microphone (Sennheiser Co., Germany), with a sampling rate of 44.1 kHz and a sampling accuracy of 16 bits. In the study area, male cuckoos regularly call when perching on electrical wires (Li et al. 2016), which enabled us to approach within 10‒30 m of calling males and obtain the best possible recording with minimal background noise. Consequently, we were able to record vocalizations of 13 different males, three of which were individually marked (banded) before recording. The fate of the other 45 banded cuckoos was unknown. Although we did not band the other ten males, we avoided repeated sampling from the same male by travelling along roads within half a day and recording new males that were at least 2 km away from males which had been recorded. For the three banded males, we recorded each individual’s calling on two, three, and five successive days respectively.

Sound analysis

We used Avisoft-SASLab Pro software (Avisoft Bioacoustics, Berlin, Germany) to resample the recordings using an 8 kHz sampling frequency and created spectrograms with the following settings: sample size, 16 bits; Fast Fourier transform length, 256 points; Hamming window with a frame size of 100% and an overlap of 50%; frequency resolution, 31 Hz; and time resolution of 16 ms. Male cuckoo calls comprise of two elements (Fig. 1) and we manually separated each element, which is a continuous signal in the spectrogram, following the procedure used in previous studies (see Fuisz and de Kort 2007; Wei et al. 2015; Zsebők et al. 2017). We then automatically measured each element: firstly, we used Avisoft-SASLab Pro software to automatically search the maximum amplitude in each element, and then determine the start and end points of each element at approximately 16 dB lever lower than the maximum amplitude. At 16 dB lever, most measured elements were explicit and above the background noise. The following variables were measured: duration of the element (t _dur1, t _dur2); duration from the start of element to the point of maximum amplitude within that element (t _dis1, t _dis2); frequency at the start point of the element (f _sta1, f _sta2); frequency at the end point of the element (f _end1, f _end2); minimum frequency of the element (f _min1, f _mim2); maximum frequency of the element (f _max1, f _max2); frequency of the maximum amplitude within the element (f _peak1, f _peak2); time interval between the first and second element (t _int). We measured ten calls for each male from each day. For two males with less than ten calls, we measured the total number of seven calls. Original measurement data of call features can be seen in Additional file 1: Table S1.

Data analyses

We collected two sets of acoustic data. The first data set contained 94 calls from ten unbanded males, and 30 calls from the three banded males, all of which were recorded on the same day. The second data set contained only recordings from the three banded males, and consisted of 20 calls from the male recorded on two continuous days, 30 calls from the male recorded on three continuous days, and 38 calls from the male recorded on five continuous days. The consistency of call features was examined using only the second (banded male) data set, whereas all other analyses were based on the first data set (combined banded and unbanded males).

We statistically described the frequency and temporal characteristics of cuckoos’ call using the average measurements for each male. We used coefficients of variation (CV) for each variable to compare differences within (CV_w) and between (CV_a) individuals (Robisson et al. 1993). We computed CV for each individual based on all calls belonging to that individual, and then calculated the mean as CV_w. We used the average value for each individual to compute CV_a. The ratio of CV_a/CV_w is the measurement of potential individual coding (PIC) which shows the importance of each variable used in identifying individuals (Charrier et al. 2001, 2003). We determined candidate variables for identifying individuals when the variable PIC value was >1, meaning that the variable showed greater variation between individuals than within an individual.

Using the first data set (combined banded and unbanded males), we standardized 12 variables, which PIC value was greater than 1, using the formula: \( {{\left( {{\text{value}} - {\text{mean}}} \right)} \mathord{\left/ {\vphantom {{\left( {{\text{value}} - {\text{mean}}} \right)} {\text{standard deviation}}}} \right. \kern-0pt} {\text{standard deviation}}} \). Based on these 12 standardized variables, we calculated the similarity of all pairs of calls using Pearson’s R for both within individuals and between individuals. Budka et al. (2015) set a value, called a ‘threshold’ that enabled them to separate the similarity of pair calls of the same male from that of different males, as the former was generally larger than the latter. Following this method, we attempted to find a threshold for individual male cuckoo identification through trial and error. We also estimated the number of males based on the threshold value: if a call’s maximum similarity (the maximum similarity between this call and all other calls) was less than the threshold, this call was identified as being from a new male. Spectrogram cross-correlation (e.g. Xia et al. 2011) was not used due to the volume of background noise in the recordings.

To compare the similarity values calculated from both the first and the second data sets, we standardized all variables from the second data set, which only contained the calls from banded males, using the mean and standard deviation calculated from the first (unbanded and banded combined) data set: \( {{\left( {{\text{value}} - {\text{mean}}} \right)} \mathord{\left/ {\vphantom {{\left( {{\text{value}} - {\text{mean}}} \right)} {\text{standard deviation}}}} \right. \kern-0pt} {\text{standard deviation}}} \). Then using the second data set, we calculated similarity (Pearson’s R) for each possible combination of calls from the same male in order to test the consistency of call features over time. We hypothesized that the similarity of all combinations of calls from different days within the same individual male would be larger than the similarity of all combinations of calls from different males.

We also used DFA (linear combination of variables that maximally separate the data points pertaining to different categories) based on the original data set with 15 variables. In the first data set, results from jack-knifed classifications are reported as percentages of songs correctly assigned. In jack-knifed classifications, each song was assigned to an individual on the basis of discriminant functions calculated from all songs in the data set except the one being classified. In the second data set, we used the 13 discriminant functions (corresponding to 13 males) constructed based on the first data set to classify calls recorded from different days. All analyses were performed using R v. 3.3.1 (R Core Development Team 2016).