End-to-end Audio Classification with Small Datasets – Making It Work
Maximilian Schmitt, Björn Wolfgang Schuller · 2019
Deep end-to-end learning is a promising approach for many types of audio classification tasks. However, in fields such as health care and medical diagnosis, training data can be scarce, which makes training a neural network from the raw waveform to the target a challenge. In this work, we focus on a public dataset of human snore sounds, categorised into four classes, where one particular class has only a few training samples. We emphasise the pitfalls that need to be taken into account when working with such data and propose an end-to-end model providing a performance similar to that of other deep and non-deep approaches. Furthermore, we show that a model using only convolutional layers outperforms a model employing also recurrent layers.