Bird Song Classification
-
© Treimer
We often hear song birds when we spend time outside in our garden, in parks or just on our way to the grocery store and often have no idea what kind of bird it is we are hearing. This project was made to help those who wonder, but cannot see the birds that are singing.
Most people carry their smartphone with them every day, a powerful device, that could potentially help the curious ones identify song birds. A quick look at the app store however shows that the bird classification apps that exist are meant to identify birds based on photos — not helpful when the bird is hidden or simply too far away — and so far only one available app can classify birds by sound. The issue with this app is, that users need to download additional packages for the birds in the region the user is in. This project tackles this problem with a web based solution.
In order to do this, a CNN was trained to identify the ten most common song birds in Germany. The songs used to train the model were first transformed into mel spectrograms and used for training, like any other image would be used for an image classification task. Then a web application that allows users to upload mp3 files was implemented. The user is then shown the most probable class for the file alongside the probability for the class and an image of the bird. To make the classification more accessible on iPhones, a WebView-App for iOS, showing the web app, was developed.
The fact that the classification is done server-side and the app essentially only displays a web page keeps the amount of data the users have to download to a minimum. More birds or regions can be added to the app without it requiring an update on the end user's devices.
