Bird call detection using deep learning

In this thesis, we compare different deep learning methods for bird sound detection. For this purpose, by using digital signal processing methods, our audio data sets containing recordings from multiple fields are turned into features as mel spectrogram images, mel frequency cepstral coefficients (MFCC) or gammatone frequency cepstral coefficients (GTCC). For our convolutional neural network (CNN), we use different layers such as input layer, convolution layer, normalization layer, activation layer, pooling layer, fully connected layer and classification layer. The gray scale mel spectrogram images are used to train our CNN for different parameter settings such as layer sizes, layer numbers, input sizes and training options. On the other hand, extracted gammatone frequency cepstral coefficients and mel frequency cepstral coefficients are used as features for recurrent neural network (RNN) based bidirectional and unidirectional long short term memory networks (LSTM). Both MFCC and GTCC are also used as input for a simple neural network algorithm. For both of our long short term memory networks, we use different number of LSTM for comparison. Accuracy of the detection is validated for all methods for different parameters using area under curve (AUC) of receiver operating characteristics.