Machine Learning for Bird Song Learning

Lead Research Organisation: Royal Holloway, University of London
Department Name: Psychology


Songbirds, including familiar species like chaffinches and great tits, share an unusual ability with us: vocal learning. Like us, birds need to hear and imitate others in order to develop their vocal communication signals. Most mammal and vertebrate species cannot do this, including all other primate species apart from us. In recent years, research into the development, neurobiology, and genetics of song learning have revealed ever deeper links between human speech and bird song - so much so that bird song currently represents the best animal model we have for understanding the biology of speech.

In order to study bird song, researchers need to accurately measure how different songs are from each other. These measures are needed to assess whether one bird really did imitate another, and how precisely they did so. Developing computer algorithms to make such measurements is difficult, however, for many of the same reasons that speech recognition is a difficult task for computers. In this grant, we will use a new approach to solve this problem - inspired by developments in speech recognition. First we will train birds to peck on buttons to get a food reward from a bird feeder, and then train them further to discriminate between different "notes" within bird songs. Then we will train "machine learning" computer algorithms to replicate the birds' decisions. We will thus develop a computer algorithm that we can use to compare bird songs in a way that is biologically validated.

We will then use our algorithm to investigate how birds learn their songs. To do this, we will make use of data-sets where researchers have simply recorded the different songs sung by birds within the population. This data contains a signature of how the birds actually learned their songs in much the same way that our genomes contain signatures of our evolutionary history. We will exploit this by using a statistical technique in combination with simulation models to infer how birds learn their songs: how frequently they generate new song types due to errors or innovations; who they prefer to learn from; and which songs they prefer to learn. We will do this for 15 different species and populations, allowing us to compare how different groups learn their songs for the first time.

Technical Summary

Bird song learning research has been built on our ability to judge the similarity between song syllables, but current methods have not been validated against birds' own perception. In order to carry out the next generation of studies of song learning, we need to develop more accurate methods, rooted in biology. And to do that, we first need comprehensive data-sets of how birds themselves perceive differences in song syllables.

Objective 1: Generate data-sets for how birds perceive differences between song syllables using operant conditioning methods, using an AXB task, for three unrelated species: zebra finch, great tit and jackdaw. We will generate around 150,000 trials.

Objective 2: Develop and train machine learning algorithms to measure song syllable similarity. Recent developments in machine learning provide powerful methods for fitting algorithms to complex time series data, like bird song syllables. We will develop and train algorithms using the results from Objective 1. We will compare the performance of our algorithm against current methods, and will host a data tournament for the machine learning field to further search for optimal solutions.

Objective 3: Apply the machine learning algorithms developed in Objective 2 to a fundamental problem in bird song learning: we lack quantitative estimates for how precisely birds learn songs. Without this information, it is impossible to take advantage of the diversity of bird song learning styles in different species and gain a comparative understanding of how song learning behaviour evolves. For this objective, we will (a) collate patterns of song sharing in populations of birds of 15 different taxa; (b) compare syllable structure of all songs within each of the populations using our algorithm; (c) use Approximate Bayesian Computing to fit the results to cultural evolutionary simulations, and thus estimate underlying parameters of learning - in particular the precision of syllable imitation.

Planned Impact

We will generate a state-of-the-art method for comparing the similarity of bird songs, and a data-set for other researchers to use when developing their own methods. Our method will be incorporated into a song-analysis program (Luscinia) that will be readily useable by members of the research field. Research that will benefit from these methods has the following impacts:
(a) Biomonitoring. Bird song is often the best record that we have of avian biodiversity - especially in tropical forests where biodiversity is highest and visibility of birds very limited. Processing hours of song recordings manually is a difficult and skilled task, and recently, interest has grown in computational methods that can automate the task. Our project will add to this by developing the first method validated by avian perception itself. Both R-Co-I Stowell (developer of Warblr), and PI Lachlan (developer of Luscinia) have a proven track record in implementing computational bioacoustic techniques for a broader audience.
(b) Biodiversity. Song often provides one of the critical phenotypic cues needed to identify new species. In some cases, song is the only clear and unambiguous character. To use song features to distinguish taxa, an accurate way to quantitatively compare songs is required; we will create and make this available to the field via the Luscinia software. The less sophisticated measures already implemented in Luscinia have already been used for this purpose, helping to identify the Gran Canarian Blue Chaffinch as a separate species from the Tenerife Blue Chaffinch, and in so doing, discovering the rarest, and one of the most endangered bird species in the E.U. Other labs are currently carrying out similar studies in Colombia and Tanzania amongst other places.
(c) Bird song neuroscience. Bird song is an established model system for speech, at a neurobiological and genomic level. Genes involved with bird song learning have been implicated in human disease. Research into this field requires accurate assessments of song structure and song similarity, which we will deliver. Through PI Lachlan's work on Luscinia, and co-PI Clayton's senior position in the bird song neurobiology field, we again have a clear plan of how we will make our methods available to a broader field and advertise them.


10 25 50

Related Projects

Project Reference Relationship Related To Start End Award Value
BB/R008736/1 04/06/2018 31/03/2019 £535,796
BB/R008736/2 Transfer BB/R008736/1 01/04/2019 03/06/2021 £426,688
Description We have been able to collect a large database on how zebra finches perceive sounds. Specifically, we have collected 20,000 trials of data about whether a bird perceives one sound to be more similar to Sound A or Sound B. To do this we used a novel "robotic bird feeder" that allowed us to work with animals in an aviary setting, with little disturbance to their normal life. We have augmented this by collecting 5,000 similar trials with captive great tits. We have also developed a machine learning algorithm that uses this data to mimic bird song perception: we can put in two novel sounds, and the algorithm will report how differently we would predict a bird would judge them to be.
As part of another objective in the grant, we have developed simulation models to explore how animal vocalisations culturally evolve. We have applied this to humpback whale song. Our simulations are able, for the first time, to explain why "revolutionary" patterns of song change occur in southern hemisphere populations, while they don't in the northern hemisphere.
Exploitation Route Our work will be used in software programmes that are used to compare bird songs. In particular, Luscinia ( is already being shaped by our findings. This software is used by a large number of researchers around the world investigating bioacoustics from ecological through to neurobiological perspectives. It will place the scientific findings from these studies on a firmer footing, and should also improve the quality of data produced. Some of these research fields are highly applied. For example, bioacoustics is widely used in biomonitoring studies.
Our work on humpback whale song helps deepen understanding of communication and change in an important species recovering from near extinction. Notably, our results show how changes in population size since the whaling-induced population crash are related to patterns of singing.
Sectors Digital/Communication/Information Technologies (including Software),Environment,Other

Description The PI contributed to a workshop organised by National Geographic (Feb 26th-27th 2019) to discuss how animal culture can be communicated to the general public, and in particular, how animal culture can be used to shape the public's understanding of the value of animal populations. This follows on from an article published in National Geographic by the PI in July 2018 on the basis of research methods developed by the PI and that are central to this grant. These methods allow us to track how animal traditions change over time. The R-Co-I has presented work from this grant in the Soapbox Science series. Work on the grant has led to the successful development of an operant device that improves animal welfare by allowing lengthy behaviour experiments to be carried out by animals living in normal group environments as well as in the wild. We have already produced a benefit in our own work - collecting data with minimal welfare impact, that previously would have involved significant time in isolation. But after communicating about our work in several impact activities, we have also shared expertise and have already supplied versions of our device to other labs around the world. We have further developed this approach by modifying our devices to work outdoors with free-living wild animals. This will potentially allow perceptual studies to be carried out without requiring animals to be kept in captivity at all. We have again received interest from other groups about our technology, and hope to supply devices and plans for building devices to other labs. This will lead to a considerable impact on animal welfare within the research field.
First Year Of Impact 2019
Sector Environment,Other
Impact Types Cultural,Societal

Title Operant bird feeder 
Description We have further refined an operant bird feeder and have successfully used it in our experiments. The new device provides several benefits from a 3R's perspective. Typically behavioural experiments on learning require animals to be kept in captivity, in isolation. Our new method obviates those needs. It uses PIT tag technology to identify individuals, and low-power/low-cost raspberry pi computers to control the experiment. Sound stimuli are presented, RFID antennae detect individuals and behavioural choices, and a motorised bird feeder controls rewards. Our device functions well without any negative stimuli. We have used our devices in our experiments for animals in aviary contexts, with wild bird outdoors. Our choices for making this device have meant that it is relatively low-cost (<£300 in components per device), allowing it to be widely used. 
Type Of Material Improvements to research infrastructure 
Year Produced 2019 
Provided To Others? Yes  
Impact We have been able to conduct our experiment without placing animals (zebra finches, great tits) in isolation. This represents a significant welfare refinement since the time taken for the experiment is considerable (>6 months). We have reached out and have already supplied our devices to other research teams in the field. 
Description Soapbox Science talk 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach National
Primary Audience Public/other audiences
Results and Impact A short talk was given online in the soapbox science series by Lies Zandberg, about the work carried out in the grant.
Year(s) Of Engagement Activity 2020