Simultaneous Species Classification and Distance Estimation for Bird Sounds Using Single-Microphone Recordings

The present work shows that it is possible to train a two-head network that both identifies the species of a bird and estimates how far it is from the recorder. This combines a classification task and a regression task in a single network. Since typically bird sound databases do not contain examples how birds sound at different distances, the key element in training a classifier is to augment the data to represent cases where birds vocalize at known distances. Here this is done by utilizing impulse responses measured in a real environment. Distance effects are added to the training data by convolving the original bird sounds with the distance-specific impulse responses. A convolutional neural network is then trained with the resulting data. This paper shows a proof of concept how the method works and gives promising results in simultaneous species classification and distance estimation using audio data from single-microphone recordings. Two cases are considered, training a network from scratch and using embeddings from an existing classifier.