Birds are one of nature's joys. Many of us engage in some form of bird watching, from feeding garden birds to visiting our island safe havens. Yet the dire survival status of many endemic New Zealand birds is well-known, with the campaigns to save the kakapo and kiwi being only the tip of the iceberg. The result is that management of bird species is required in order to ensure their survival, which includes monitoring to detect population changes over time.
One challenge with monitoring birds is that they are hard to spot even when the watchers are in the correct place, and so it is hard to estimate numbers accurately. One solution to this is to use bird vocalisation, which then means that automatic recorders can be left in the bush for weeks or months, and the resulting recording analysed to recognise types of bird, and estimate counts. Unfortunately, while the recording is automatic, the analysis is anything but. It takes an expert at least one hour to examine ten hours of recording using sonograms, a daunting task when many recordings are often collected simultaneously!
On a more individual level, bird lovers on hearing birds would like to be able to identify them. The rise of smartphone technology has enabled apps that perform this task for music, and the appeal of such an app for birdsong is clear. The challenges of both these tasks are similar: the recordings are potentially noisy (with wind, rain, and other sounds added), the birds are at a variable distance and angle from the microphone, and the precise song of birds varies immensely within a species.
We have started to develop an approach to birdsong recognition using a variety of machine learning and sound analysis approaches. The initial results are promising, but a lot more development work is required. Further we do field experiments uncovering the effect of environmental factors such as wind and rain on recordings because the performance of recogniser is heavily depends on the quality of recordings provided.
- Birds won’t sing nicely straight into the microphone
- Field recordings are susceptible to diverse noise (other animals, environmental noise, human-related noise, and recording error)
- Birds overlap their calls
- Birds have large repertoires
- Dialects (geographic variation)
We have already done most of feature representation and pre-processing work while we are currently experimenting call extraction and feature matching based on denoising results.
- Wavelet based denoising
- Machine learning, signal processing, and feature extraction
- Field experiments to reveal the effect of environmental effect on birdsong recording
We have devised a method to reduce noise and recover birdsongs embedded in noisy recordings.
Some examples showing initial and denoised versions. Clickfor more examples and source code.
We have tested different feature representation methods including Mel Frequency Cepstral Coefficients (MFCC), Thin Plate Splines (TPS), and Linear Predictive Coding (LPC) while Artificial Neural Network (ANN) for classification phase. The results were encouraging and we did a poster presentation at Higher Education for the 21st Century (HETC) Symposium 2014 held 7-8th July 2014, Colombo, Sri Lanka.
Knowing the environmental effects on birdsong recordings is also important when estimating bird populations based on field recordings, but not enough information is available in the literature. Many people including ecologists, wildlife managers and researchers use acoustic recorders in the field, but nobody is sure that the recordings they collect will be actually useful. For example, wind noise is a prominent obstacle in New Zealand. But no one knows what the maximum wind that we can still collect useful data is. Not only in automated recognition but also in manual processing the quality of recordings is very important. Therefore understanding the effect of these factors would certainly be useful. Therefore, we conducted a series of field experiments to uncover the effect of environmental factors on birdsong recording. Here we choose to playback bird songs and re-capture them at different distances and angles. The speaker volume was set to the true amplitude of the bird as far as possible because we are interested to know what the maximum distance a song can travel and what the maximum distance a song can be recognised by our proposed recogniser. Basically the purpose was to analyse the effect of wind, re-recording distance, transmission height on bird song recordings under two different habitats (open area and dense forest) during day and night. We have completed the experiment and currently are analysing the data.
The next stages of this research leading to the full development of a recogniser are:
- Convert the current research focused Matlab implementation of the birdsong denoising method into a compiled language like C in order to reduce the high computational cost.
- Use weather forecasting (or weather records during the recordings) to schedule the recorders (or to automatically discard useless (noisy) recordings), or build in a noise estimation module that decides when the noise is too high.
- Almost all current population estimations are conducted during the breeding season due to the cost and the difficulty/ infeasibility of manually analysing the long recordings. While it is useful to estimate low populations, vocal activity monitoring during non-breading season would be interesting, useful, and probably would reveal hidden information.
- Once we quantify birdsong the next question is what the actual population equivalent to this is. Therefore developing a protocol that helps to convert the number of calls found in a given time period into the number of birds with high accuracy combining the knowledge of biologists, ecologists and statisticians and engineers. This will vary between bird species and also depend on the environment where they live, amongst other factors.