Cultures in humans and other species are maintained through interactions among conspecifics. Declines in population density could be exacerbated by culture loss, thereby linking culture to conservation. We combined historical recordings, citizen science and breeding data to assess the impact of severe population decline on song culture, song complexity and individual fitness in critically endangered regent honeyeaters (Anthochaera phrygia). Song production in remaining wild males varied dramatically, with 27 % singing songs that differed from the regional cultural norm. Occurring at low population density 12% of males completely failed to sing any species-specific songs and instead sang other species’ songs. Atypical song production was associated with reduced individual fitness, as males singing atypical songs were less likely to pair and nest than males that sang the regional cultural norm. Songs of captive-bred birds differed from those of all wild birds. The complexity of regent honeyeater songs has also declined over recent decades. We therefore provide rare evidence that a severe decline in population density is associated with loss of vocal culture in a wild animal, with concomitant fitness costs for remaining individuals. Loss of culture may be a precursor to extinction in declining populations that learn selected behaviours from conspecifics, and therefore provide a useful conservation indicator.

(a) Data collection

We used data from all regent honeyeater sightings throughout the species’ range from July 2015 to December 2019 to estimate the distribution and density of the remaining wild population. Regent honeyeaters can be sexed in the field based on a combination of their size, plumage traits, behaviour, vocal attributes and in the hand during marking with unique combinations of coloured leg bands²⁵. The database consisted of confirmed public sightings reported to BirdLife Australia and data from a standardized national monitoring program based on 1367 sites throughout the breeding range²⁵. We identified males individually through a combination of colour bands on the focal male or partner female (n = 93), nest location (n = 68), unique song attributes (n = 21) or a lack of other males nearby (n = 42)²⁵. We recorded males’ songs using a Sennheiser ME62/K6 microphone on a Telinga parabola and a Marantz PMD661 digital handheld recorder. We recorded captive-bred birds either shortly after their release into the wild in 2017 or in captivity in August 2019 and obtained historical song recordings (date 1986 - 2012) from the Atlas of Living Australia and private sound collections. All individuals included in this study were at least 1 year old. See Supplementary Methods for further details of the captive breeding program and the historical song recordings.

(b) Song classification

Wild male regent honeyeaters typically produce three distinct vocalisations: a soft, ‘mewing’ call; an alarm call consisting of a squawk and/or monosyllabic squeak; and a highly distinctive song, consisting of sub-chatter building to a crescendo of a guttural warble  produced with characteristic head-bobbing (Supplementary files S1 & S2).

The sightings database included 228 wild males identified since standardized contemporary monitoring commenced in 2015. We noted the songs of 146 of these males and were able to obtain quality recordings, defined as a high signal to noise ratio and minimal background noises, of the songs of 47 of them. These males occurred in two different geographical regions, the Blue Mountains south of latitude -31.55 and the Northern Tablelands north of latitude -30.45 (Figure 1A). We tested whether species-specific song types produced in these regions were significantly different using stepwise discriminant function analysis (DFA) of 15 song parameters (see Table S1 and data analysis section below). Eighteen of the 146 males, located throughout the contemporary range, failed to sing any species-specific songs and instead produced the song of a different bird species (Figures 1A and 2). We classified these birds as ‘interspecific singers,’ based either on visual similarities between spectrograms of the songs of interspecific singers and of the species whose songs they had learned (n = 8) or the memory of an experienced observer (n = 10). We also obtained song recordings of 12 captive-reared males (three recorded one week post-release in 2017 and nine recorded in captivity in 2019) and historic recordings of 14 wild males that were recorded prior to 2012 in the Blue Mountains. We tested for differences in the songs of the males of these five categories (Blue Mountains, Northern Tablelands, interspecific, captive-reared and historic) using the same DFA procedure described above. A single observer with 6 years’ experience of monitoring regent honeyeaters (RC) obtained all but seven contemporary song recordings.

(c) Data analysis

For all data analysis, we used R v3.4.3³⁰ unless otherwise stated. We tested for spatial autocorrelation in the song type of contemporary wild males with correlograms of Moran's I, using package ncf v1.2–5³¹. For acoustic analysis we first used Audacity v2.2.2³² to isolate songs in sound files and reduce background noise. We imported the trimmed .wav files of sufficient quality (n = 73 including contemporary wild, historic wild and contemporary captive birds) into warbleR v1.1.22³³. After restricting the frequency range to 0.5 – 5 kHz, we manually selected the start and end coordinates of each song, used sig2noise to increase the signal to noise ratio (type = 3) and trackfreqs to identify the spectral components of each spectrogram. We visually inspected spectrograms to ensure track frequencies selections were representative of the spectral components of each song and used function specan to quantify 20 spectral attributes of each song (Table S1). We manually calculated a further three attributes based on visual and audial inspection of the recordings¹¹: number of syllables, number of unique syllables and number of notes per syllable. We checked for pairwise correlation across all attributes using ‘GGally’ v1.4.0³⁴, but no attributes showed consistent strong correlation (R > 0.5 or < -0.5). We log-transformed modulation index, kurtosis, maximum dominant frequency and notes per syllable to fulfil normality assumptions.

We used JMP version 15.0 to conduct a discriminant function analysis (DFA) of songs by song type. We only included significant acoustic attributes (n = 15) in the final model via a backwards stepwise selection procedure (Table S1), and assessed the fit of the model by calculating the proportion of songs assigned to the correct song type.

To determine whether interspecific singers more frequently occurred at lower population density than species-specific singers, we calculated for each wild male the number of other males sighted in the same breeding season (June to January) within distance bands of <1 and <50 km. We considered these 2 spatio-temporal categories of ecological relevance to song learning, given the regent honeyeater’s range size and capacity to undertake long-distance movements²³. We used Mann-Whitney U tests to look for a difference in the number of conspecifics located within both spatio-temporal windows, with interspecific singer or not as the binomial response.

To quantify differences in song complexity between song types, we took 14 of the acoustic attributes that represented attributes of song complexity (Table S1) and fitted a general linear model of each attribute by song type using lme4 v1.1-21³⁵.

To assess the fitness costs of males’ songs in the remaining wild population, we used logistic regression models with a binomial error structure and logit link function in lme4. For fitness analyses, we included in the dataset males whose songs were either not recorded or were not recorded of sufficient quality for acoustic analysis, but could be assigned with high confidence to a song type in the field (n = 105) because they were clearly heard singing by an experienced observer (RC). We classified the songs of a further 63 males, whose songs we could not assign to a song type as ‘unknown’ because we did not hear or record these males singing at the time they were detected, and not because their songs were intermediate between song types. The first model tested the effect of song type on whether a male was paired to a female or not. The second model tested whether song type affected the probability of paired males reaching the egg stage of nesting. The third model tested whether song type affected the probability of nesting males successfully fledging young. We then re-ran each model, reclassifying each male’s song type binomially as ‘regional cultural norm’ or ‘non regional cultural norm.’ We defined regional cultural norm as the typical Blue Mountains song in the Blue Mountains, and the Northern Tablelands song in the Northern Tablelands. We considered all other classified songs in each breeding area as ‘non regional cultural norm.’

To confirm that any fitness costs of male song were associated with differences from the regional cultural norm and were not an artefact of song type classifications, we repeated the ‘paired’ and ‘nested’ logistic regression analyses, replacing the song type of each male located in the Blue Mountains for which we had a high quality recording (n = 34) with the Mahalanobis distance of each males’ song from the multivariate mean of the entire Blue Mountains population. We calculated the Mahalanobis distance of each male’s song using heplots v1.3-5³⁶. Larger Mahalanobis distances represent greater song divergence from multivariate mean³⁷. For this analysis, we defined the regional cultural norm as the multivariate mean, rather than the most common song type. The small sample of quality recordings from the Northern Tablelands (n = 7) precluded us repeating logistic regressions with Mahalanobis distances on the Northern Tablelands population.

To assess song repeatability, we used a Mantel test in ade4 v1.7-15³⁸ with 9999 permutations to compare the song similarity distance between repeat recordings of the same individuals to the average distance between all other males’ songs. Metadata detailing how we identified individuals and which individuals we included in each component of the statistical analysis is provided in the Supplementary Data.

REFERENCES:

11. Paxton K L, Sebastián-González E, Hite J M, Crampton L H, Kuhn D, Hart P J. 2019 Loss of cultural song diversity and the convergence of songs in a declining Hawaiian forest bird community. R. Soc. Open Sci. 6, e190719.

23. Franklin D C, Menkhorst P W, Robinson J L. 1989 Ecology of the regent honeyeater Xanthomyza phrygia. Emu 89, 140-154.

25. Crates R A, Rayner L, Stojanovic D, Webb M H, Terauds A, Heinsohn R. 2019 Contemporary breeding biology of critically endangered regent honeyeaters: implications for conservation. Ibis 161, 521-532.

31. Bjornstad O N. 2018 Package ‘ncf’ v1.2-5: spatial covariance functions. https://cran.r- project.org/web/packages/ncf/ncf.pdf.

32. Team Audacity. 2014 Audacity (r): Free audio editor and recorder [computer program].

33. Araya-Salas M, Smith-Vidaurre G. 2017 warbleR: an r package to streamline analysis of animal acoustic signals. Methods Ecol. Evol. 8, 184-191.

34. Schloerke B. 2018 Package GGally. Extension to ‘ggplot2.’ https://cran.r-project.org/web/packages/GGally/GGally.pdf.

35. Bates D. 2019 Package ‘lme4’ v1.1-21: Linear mixed-effects models using ‘Eigen’ and S4. https://cran.r-project.org/web/packages/lme4/lme4.pdf.

36. Fox J, Friendly M, Monette G. 2018 Package ‘heplots’: visualising hypothesis tests in multivariate linear models. https://cran.r-project.org/web/packages/heplots/heplots.pdf.

37. De Maesschalck R, Jouan-Rimbaud D, Massart D. 2000 The Mahalanobis distance. Chemomet. Intel. Lab. Syst. 50, 1-18.

38. Dray S, Dufour A-B, Thiolouse J. 2020 Package ade4: Analysis of Ecological Data: Exploratory and Euclidean Methods in Environmental Sciences. https://cran.r-project.org/web/packages/ade4/ade4.pdf.

The dataset contains X files:

• "Recordings_summary.xlsx" = full details of each male regent honeyeater, where, when and how it was recorded, whether that male was observed and or recorded in multiple years, which analysis each male could be included in and which sound file(s) correspond to which male. We have withheld sensitive location data (lat,long) because (i) of the critically endangered population status of the species and (ii) it is not required to replicate the data analysis with the files available here. Specific location data is available from the corresponding author upon reasonable written request.
• "Recordings for manuscript analysis.zip"  = .wav regent honeyeater song sound files used in the manuscript.
• "Repeat recordings.zip" = .wav regent honeyeater song sound files, including repeat files of some males where available, for use in modelling song repeatability.
• "Regent honeyeater song script.R" 
• "paired1.csv" = for re-running GLM of pairing status of males by song type and cultural norm.
• "nested2.csv" = for re-running GLM of nesting status of paired males by song type and cultural norm.
• "fledged1.csv" = for re-running GLM of nest success of nesting males by song type and cultural norm.
• "complexity_conspecs_mahal.csv" = for calculating Mahalanobis distances for the Blue Mountains population.
• "mahala.paired.bmtn.csv" = for re-running logistic regression of pairing status of Blue Mountains males by Mahalanobis distance.
• "mahala.nested.bmtn.csv" = for re-running logistic regression of nesting status of paired Blue Mountains males by Mahalanobis distance.
• "conspecifics.csv" = for re-running Mann-Whitney U tests on conspecific density versus species-specific or interspecific songs in the wild population.
• "manualoc_output1.csv" = for specifying start and end timings of songs in sound files for analysis in warbleR.
• "regent_PCA_params3.csv" = for multivariate analysis of spectral components of regent honeyeater songs- checking cross-correlations, distributions, log-transformations etc.
• "complexitymetrics1.csv" = for GLMs of song complexity by song type.
• "manualoc_duplicates.csv" = for specifying start and end timings of repeat recordings for analysis in warbleR.