This paper discusses distinguishing characteristics of the Xenocanto bird sound collection. The main aim is to indicate the relation between automated recognition of bird sounds (or feature recognition in digital recordings more generally) and curating large bioacoustics collections. Not only do large collections make it easier to design robust algorithmic approaches to automated species classi ers, those same algorithms should also become useful in determining the actual content of the collections.
For the past two years the BirdCLEF challenge [
Sjoerd Mayer's \Birds of Bolivia" CD-ROM's [
The authors essentially took these concepts a step further, and designed and constructed an interface to a non-commercial, open database situated on the world wide web. A number of guiding principles were formulated that distinguished XC from other sound collections at the time: { Anyone with web access is invited to upload sounds. XC does not refuse recordings. Contributors can share any bird sound they nd interesting, provided they are below a xed maximum size (initially 1 MB, now 10 MB) and provided a required minimum set of metadata is given: species, recordist name, location name, country, recording date, time of day, elevation, and sound type(s). This system certainly has drawbacks: a considerable fraction of the recordings is short, of dodgy quality, or both. Still, such recordings may be useful. They may represent poorly known locations or vocalisations, or may simply contribute to the sample size of individual species. Also, in the context of automated species identi cation algorithms, it is clear that any real-life deployment of such an algorithm would have to deal with poor quality recordings as well. { The recordings uploaded to XC are shared. Re-use of the recordings is intended, for purposes that are in line with the aims of XC, such as downloading to personal collections, embedding sounds in educational or personal web sites, use for scienti c research, etcetera. The Creative Commons licenses (http://creativecommons.org/) o er a useful framework. After consultation with the community it was decided to settle for CC-BY-ND-NC (attribution, no-derivatives, non-commercial) licenses. Since this is in fact a rather restrictive license, nowadays one can also choose CC-BY-NC-SA (SA stands for share-alike) and CC-BY-SA licenses that allow more liberal re-use. In all cases attribution of the author/contributor on republication is mandatory. For discussion of the limits of the other terms, see the Creative Commons website. The XC website code is written in free, open source software. It is based on a standard LAMP (Linux, Apache, MySQL, PHP) set-up, with some additional software written to show sonograms, implement mapping, and so on. { Anyone can contribute to the collection in some way. Apart from sharing recordings, people may contribute expertise on identi cation, set identi cation challenges, o er experience with equipment, write articles on-site, or just comment on recording achievements. { Anyone can challenge an identi cation (ID) on the site. The vast majority of recordings have been identi ed correctly to species by the recordist, but errors are inevitable. When challenged, the recording is set aside and does not appear in search results until the ID is resolved by the community. This is usually done in an open discussion on the forum. If the ID is agreed upon, the recording is put back into the collection by the administrators. The administrators therefore have the role of arbiters, rather than authorities, and in fact there are no designated authorities that decide on species identi cation. This is one of the more uncommon features of Xeno-canto, and in this sense it di ers from other well-known community projects on natural history, such as eBird (ebird.org), Observado (waarneming.nl / observado.org).
At present, May 2015, the XC collection contains some 243,000 recordings from over 9,300 bird species, shared by more than 2400 contributors from all over the world. In the rest of this paper, the development and current status of the XC collection are illustrated and a few points relevant to its relation with automatic sound classi cation and recognition are discussed.
The growth of XC is illustrated in Figures 1 and 2, by plotting the number of recordings and the number of contributors over time. Two things are noteworthy. Firstly, the data for the initial period is incomplete, since the uploading dates were initially not recorded. Secondly, there are pronounced seasonal e ects, most obvious in the number of contributors. These points are remedied to some extent in subsequent gures by plotting versus the number of recordings instead of versus time. 2.1
Both the number of recordings and the number of contributors grow at increasing rates. Remarkably, plotting the number of recordings versus the number of contributors shows that they have consistently increased at approximately the same rate. See Figure 3. This leads to a more or less constant average number of recordings per contributor, which turns out to be about 100. However, it should be noted that the distribution of recordings per contributor is very broad and skewed. At this moment 298 contributors contribute more than 100 recordings, and many more contributors, around 2100, contribute less than 100 recordings.
The three largest contributions each comprise more than 10000 recordings, more than 100 times the average; 729 contributors contributed 1 recording, 100 times less than the average. The Zipf-like plots in Figure 4 serve to characterise the distribution at various stages during the development of XC. 2.2
When a sound is uploaded a set of metadata is required, among which the name
of the species. Specifying the subspecies is optional. The taxonomy of the site
was initially based on the taxonomy in Neotropical Birds [
IOC 4.1 recognises 10,518 extant species and 150 extinct species; to this list XC has added 16 additional recently described or as yet undescribed species. About 9330 are represented in XC at this moment. To our best knowledge this constitutes the largest number of species in any public collection of bird sounds. (There is at least one private collection that has more species, but it includes sounds of all species from XC.)
The growth of the number of species may provide a clue about the moment of completion of the collection at the species level. Figure 5 shows the species accumulation curve up to may 2015, together with a randomised accumulation curve and a t used for extrapolation. The randomised version is based on a random draw from all recordings present in XC. Clearly the two curves di er signi cantly. This is caused by the fact that XC started out with only Neotropical species, and that other world areas were added later. The randomised species accumulation curve does not take that into account. The two curves are seen to meet up after about 170000 recordings, well after XC went global.
For any number of reasons (abundance and size range, vocal (in)activity, accessibility of the range of the species, accessibility of the site to name just four) the recordings are not evenly distributed across the species. The current expectation value for the number of recordings per species is around 20. However some 20 species have attracted over 500 recordings, while around 1200 are still waiting to be uploaded. The distribution plotted in Zipf-fashion is shown in Figure 6, probability densities are shown in Figure 7.
The species abundance curves can be extrapolated into the future by making assumptions on the probability that species that are not represented at this time will be uploaded. A reasonable t is achieved by assuming that the probability of a new species being uploaded is 1/3 of that of a species with 1 recording in XC, with the ratios between probabilities of species already represented remaining equal. An extrapolation based on that assumption is shown in Figure 8. Of course the extrapolation follows the randomised species abundance curve very well. The extrapolation is shown up to 900,000 recordings, at which point it is still about 600 species shy of the total number of species. The precise number will depend on the assumptions made, but it seems reasonable to assume that completion at the species level will take a multiple of the number of recordings present at this moment.
The BirdCLEF workshop requires entrants to identify recordings from the XC collection to species level based on the species level identi cation provided by the XC community. It is worthwhile to have another look at the species level IDs in XC. For a number of reasons the ID to species level, even if correct, may be misleading.
{ Presence of unnamed background species Although recordists are asked
to mention the background species present in the recordings, not all recordists
do so. On average 2 species are identi ed per recording, but it is certain that
many more could be identi ed. Interestingly, the presence of named
background species helps humans to identify a sound of interest (as the authors
know from personal experience), but this does not seem to lead to a higher
rate of identi cation in the algorithmic identi cations [
Other contributions to the diversity of sounds which do not necessarily align with subspecies categories are geographical dialects, such as in Yellowhammer (Emberiza citrinella), and the size of the vocabulary of a species, such as in Common Nightingale (Luscinia megarhynchos ). Little quantitative information is available on the extent of dialect formation and the size of the vocabulary across the range of the overwhelming majority of the 10518 species of birds. Apparently the e ect of such diversity at the species level on the results of automatic recognition has not been quanti ed yet. Intuitively, given a set of training data, one would expect a species that shows little diversity to be recognised more faithfully than a species that shows a lot of variability.
In [
The results from the 2014 and 2015 BirdCLEF challenges o er an interesting perspective on the use of automated algorithmic techniques on the one hand, and large accessible public archives of sound data on the other.
At present, the focus in the challenges lies squarely in the eld of automated recognition, and understandably so. The large Xeno-canto database has been the basis of the challenges, and give the rst general insights in automated feature extraction and classi cation to species level for general vocalizations. The species set included in the latest 2015 edition spans 1000 species with a huge range of di erent types of bird songs and calls. The BirdCLEF paper in this volume contributes to our understanding which techniques excel at this type of challenge.
We would welcome a second application of the algorithms, however, one that would allow a deeper insight into the variety of vocalizations actually represented in archives such as Xeno-canto. There is great potential for collaborative projects, in which estimates would be computed of a number of statistics. Example include (a) estimates of repertoire sizes in song birds (or other taxa); (b) discovery of subspecies with di erent vocal signatures; (c) the ability to extract a small representative sample of di erent vocalizations for focal species, or focal localities. We hope that we may attract the computer science community to work with us to start to address these types of challenges.