Contribution
Investigation of binaural distance estimation with artificial neural networks trained on simulated data models
* Presenting author
Abstract:
It is well-known that different cues (such as intensity, DRR, timbre, binaural differences) are important for auditory distance perception in humans. However, it remains unclear which specific signal features are essential and most successful in binaural distance estimation for machine learning algorithms. Since the amount of real-world recorded and labeled data is limited and highly depending on particular room acoustic conditions as well as source and listener configurations, auditory virtual environments (AVE) can be valuable for simulating new and diverse data. However, the extent to which simulated data represents real data may be highly variable, raising questions about how well distance estimation models trained on synthetic data will perform with real-world data. To address this modeling gap, we investigate different model approximations in an AVE for synthetic data generation. We then study the robustness of different feature combinations in conjunction with a recurrent neural network, the influence of different AVE models for the generation of training data, and examine the impact of these factors on distance estimation in a real environment.