Bird Audio Recognition Using Convolutional Neural Networks

With modern-day technological advances, creating and developing innovations that greatly improve the standards of life, it is also important to consider their environmental impact. Side-effects of these innovations are costly to the environment. A method to monitor environmental sustainability is to use birds as a bioindicator of an area. The greater variety of bird species present in an area suggests a superior habitat quality. People can take pictures or set up cameras to capture photographs of birds to record the species present. However, to capture an identifiable photo is difficult birds are erratic and sensitive to movement. In contrast, the sounds of birds travel far distances and are easily recorded. Therefore, we propose to develop a program to identify bird species using audio recordings. In this poster, we present our method to identify the species of birds through analyzing audio profiles in convolutional neural networks (CNN). The first part visualizes the audio profile, using a MatLab program. The CNN algorithm analyzes two-dimensional images instead of one-dimensional audio files. Therefore, the MatLab program applies the Short-Time Fourier Transform onto the waveform audio file to create a spectrogram. It visualizes an audio profile by representing the magnitude of the signal frequencies in relation to time. The spectrograms are cropped to four seconds segments to increase the amount of training data and unify the spectrogram dimensions. Then, the next section inputs the spectrograms into a CNN algorithm. The CNN algorithm analyzes the spectrograms to extract the distinct features present throughout the spectrograms of one species. Since each bird has a distinct vocal sound, it has a distinct pattern in the spectrogram. This two-part process is reiterated for various species to introduce variety in the database as well as test the accuracy of the algorithm. In the current stage, the CNN algorithm is trained to identify a few local species in Maryland. The algorithm requires additional fine-tuning to ensure accurate identification for every input. To create a program that promotes user convenience, integrating the components of the MatLab program and CNN algorithm into a streamlined program is in development. Moreover, once the CNN algorithm is finalized, we plan to expand the database of recognized birds by training the algorithm with more data. With this program, we aspire to encourage environmental sustainability by using birds as bioindicators. People can be a part of monitoring the welfare of their environment by recording nature’s music.