Cornell Birdsong Identification Competition

This summer, we have completed our first ever competition as Team Epoch. The Cornell Birdsong Identification Competition. This competition has given us many insights in how we can improve our way of working as a team. With all these experiences we are now looking forward to be even more competitive in the next competition.

Cornell is an American university that is one of the world’s leading organizations when it comes to applying AI. They issued the Birdsong Identification competition to help speed up the process of automatic bird call detection. This helps them to use AI to track endangered bird species, because human conservative efforts are not sufficient to keep track of all species.

The goal of the competition was to design an AI model that is able to classify whether or not a specific bird can be heard in every 5 seconds of a sound clip. Even bird experts often find this task hard, especially because the competition aims to classify 264 different species. One of the challenges was to make an algorithm that is smart enough to listen to the bird calls while ignoring the background noise. Listen for yourself, to see if you can hear any birds!

The trick was to design an architecture that can interpret the sound waves and filter out noise at the same time. This was accomplished by the DenseNet model. This model consists of convolutional and recurrent units in series. The convolutional units allow the model to focus only on the necessary bird sounds and the recurrent units allow the model to remember previous parts of the sounds.

Since this was our first competition, a lot needed to be figured out and we finished in the top 75%. However, our results improved drastically in our second competition, so check it out on the Lyft Autonomous Vehicle page!