Few-Shot Sound Source Distance Estimation Using Relation Networks

In this paper, we study the performance of few-shot learning, specifically meta learning empowered few-shot relation networks, over classic supervised learning in the problem of sound source distance estimation(SSDE). In previous research on deep supervised SSDE, obtaining low accuracies due to the mismatch between the training data(sound from known environments) and the test data(sound from unknown environments) has almost always been the case. By performing comparative experiments on a sufficient amount of data, we show that the few-shot relation network outperform a classic CNN which is a supervised deep learning approach, and hence it is possible to calibrate a microphone-equipped system, with a few labeled examples of audio recorded in a particular unknown environment to adjust and generalize our classifier to the possible input data and gain higher accuracies.