Normalization of audio signals for the needs of machine learning

Trends in machine learning development increasingly indicate that the main limitations of a model are the availability of labeled and the high quality data for model training. On the other hand, for general experiments, there has never been more data available on the internet that can be processed and utilized in machine learning. Therefore, it is essential to investigate methods for enhancing and expanding existing training datasets. In this paper normalization of audio signals for the needs of machine learning is analyzed. Here, the focus is on loudness normalization and possible effects valuable in modern approaches. The results show that loudness based normalization may affect morphological characteristics of an audio signal in temporal domain.