danoynay / bird-sound-detector-using-deeplearning-techniques Goto Github PK

The island of Cebu is known to have an abundant avifauna biodiversity in the Philippines. Cebu is habitat to about 270 species of birds, 9 of which are endemic birds to the Philippines. Birds have been regarded as hallmark of biodiversity. Previous studies prove that fluctuations in bird population reflects changes in the environment. Aiming to promote the conservation of the avifauna’s biodiversity, this study created methodological classifications of selected bird species found in Cebu. This is done through machine learning of the audio-data of bird sounds that are converted to MFCCs (Mel Frequency Cepstral Coefficients) using deep learning technique, specifically CNN (Convolutional Neural Network) and LSTM (Long Short-Term Memory) - RNN (Recurrent Neural Network). Methods include pre-processing, training, validating and testing which revealed validation accuracy of 96.24% and 95.57% for CNN model and LSTM model, respectively. In addition, CNN model garnered 88% of prediction accuracy of separated audio-data comparing to 85.5% using LSTM model. CNN also got the fastest running time than LSTM in model training and validation process yet, LSTM model has the smaller memory size comparing to CNN model. Comparatively, the study revealed that CNN trained model is more reliable in classifying and predicting bird species using audio-data of bird sounds compared to LSTM- RNN model.