The growing usage of machine learning and artificial intelligence within audio signal processing underscores the significance of high-quality audio datasets for advancing audio algorithms such as speech enhancement, echo cancellation, de-reverberation, and blind room estimation. However, the conventional approaches of collecting such data present various limitations. Measurement based approaches are costly and time-consuming, and synthetic data generation using standard acoustics simulation methodology has been shown to generalize poorly to real world scenarios, due to limitations in capturing the intricacies of real world room acoustics. In this paper, we present a framework that offers a solution to the challenges associated with dataset creation, enabling the efficient production of extensive datasets that closely mimic real-world audio scenes, thereby enhancing the efficacy of machine learning models. Through the lens of a specific use-case illustration, we highlight the integration of a hybrid wave-based / geometrical acoustics simulation for dataset generation. Notably, our focus extends to accurate device modeling—a critical aspect for the development of multiple-microphone devices and subsequent refinement of machine learning algorithms. We illustrate how the dataset accuracy surpasses the standard limitations of geometrical acoustics simulations. We present analysis of the computational performance of the system and we demonstrate examples of improved machine learning performance using the data.
(See document for exact affiliation information.)
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member Join the AES. If you need to check your member status, login to the Member Portal.
