Special Issue on Model-Based and Data-Driven Audio Signal Processing

“All models are wrong, but some are useful” - understanding “models” as analytical mathematical models, this aphorism, originating from George Box in 1976, motivates the synthesis of model-based and data-driven audio signal processing as the leitmotif of this special issue.

Acknowledging that current analytical models alone cannot provide the performance and sophistication that state-of-the-art systems should be endowed with, in the last decade, we have witnessed a rapid paradigm shift from model-based algorithms to data-driven ones, using primarily deep neural networks (DNNs), with many successful solutions in diverse application areas, such as speech and audio enhancement, source separation and localization, dereverberation, sparse representations of audio signals, audio rendering, acoustic event detection, music information retrieval, and more.

However, learning-based methods, especially those based on DNNs, do not usually embrace the physical nature of the problem and rather optimize the nonlinear relationship between training data and expected results, relying on only computational power. Many problems call for more efficient solutions that minimize the computational burden, amount of training data, memory requirements, and, consequently, power consumption. Moreover, pure data-driven DNN approaches remain largely unexplainable and noninterpretable. In addition, uncertainty regarding the DNN behavior for previously unseen and dissimilar input data should be avoided in most cases.