Deep Learning For Natural Sound Classification

Nowadays, it is very common to use sensors for controlling the population of different animal species in a natural environment. A large number of sensors can be deployed in wide areas and they will capture information relentlessly, producing a huge amount of data. However, analysing the collected data by humans is a big challenge and for that reason, it is necessary to develop automated technologies in order to help experts on that task. Within this context, we present an automatic system to detect and classify sounds, especially those generated by birds and insects among other sounds that can be heard in a natural environment. For the development of the system, it has been necessary to generate a sound database. The recorded database consists of field recordings in three different Natural Parks, with sounds of several bird and insect species, as well as background noises. The automated system employs state of the art neural networks for detecting and classifying sound frames. Experiments were done using several signal preprocessing and acoustic features. The experiments show a good accuracy in detection and classification of sound frames and with results higher or comparable to other state of the art approaches.