A framework for generating large-scale microphone array data for machine learning

Abstract

AbstractThe use of machine learning for localization of sound sources from microphone array data has increased rapidly in recent years. Newly developed methods are of great value for hearing aids, speech technologies, smart home systems or engineering acoustics. The existence of openly available data is crucial for the comparability and development of new data-driven methods. However, the literature review reveals a lack of openly available datasets, especially for large microphone arrays. This contribution introduces a framework for generation of acoustic data for machine learning. It implements tools for the reproducible random sampling of virtual measurement scenarios. The framework allows computations on multiple machines, which significantly speeds up the process of data generation. Using the framework, an example of a development dataset for sound source characterization with a 64-channel array is given. A containerized environment running the simulation source code is openly available. The presented approach enables the user to calculate large datasets, to store only the features necessary for training, and to share the source code which is needed to reproduce datasets instead of sharing the data itself. This avoids the problem of distributing large datasets and enables reproducible research.