Classification of Respiration Sounds Using Deep Pre-trained Audio Embeddings

Abstract

In this work we present the use of an end-to-end deep learning based pre-trained audio embeddings generator, and apply it to the purpose of classification of respiration sounds. With this approach, there is no need to pre-compute spectral representations, e.g. MFCC or filterbanks, since the classification model uses raw audio as the input. Transfer learning was used to train an audio classifier for sounds of respiratory cycles as defined in the ICBHI 2017 challenge. The results on this dataset show that this end-to-end model represents a viable alternative to more common spectral-based classifiers, while achieving state-of-the-art performance.