New aspects in birdsong recognition utilizing the gabor transform

In this paper we will be concerned with mathematical methods for birdsong recognition and classification. Current approaches compute the spectrogram of an audio recording using the Gabor transform, which is then used as input for a convolutional neural network (CNN) to classify the recording. While recent work is dedicated to finding the best hyperparameters for training the CNN and for data augmentation, the parameters for the Gabor transform and the signal detection methods receive less attention. We aim to close this gap by evaluating the effect of different window lengths on the overall classification accuracy. Additionally we propose a method for denoising signals based on quantiles to account for different signal to noise ratios in the dataset.