Audio Explainable Artificial Intelligence: A Review

Abstract

Artificial intelligence (AI) capabilities have grown rapidly with the introduction of cutting-edge deep-model architectures and learning strategies. Explainable AI (XAI) methods aim to make the capabilities of AI models beyond accuracy interpretable by providing explanations. The explanations are mainly used to increase model transparency, debug the model, and justify the model predictions to the end user. Most current XAI methods focus on providing visual and textual explanations that are prone to being present in visual media. However, audio explanations are crucial because of their intuitiveness in audio-based tasks and higher expressiveness than other modalities in specific scenarios, such as when understanding visual explanations requires expertise. In this review, we provide an overview of XAI methods for audio in 2 categories: exploiting generic XAI methods to explain audio models, and XAI methods specialised for the interpretability of audio models. Additionally, we discuss certain open problems and highlight future directions for the development of XAI techniques for audio modeling.