Colloquium - Details
Bachelor-Presentation: Estimation of Head Rotations from Binaural Signals with Deep Learning
Michael Wiefeld
Tuesday, September 29, 2026
02:00 PM
IKS 4G | hybrid
This thesis investigates the estimation of head rotations around the yaw axis directly from binaural audio signals using deep learning. Head motion causes characteristic temporal changes in binaural cues, but the estimation task is complicated by ambiguities that can make opposite rotation directions difficult to distinguish.
A convolutional–recurrent neural network is used to estimate the relative head-rotation trajectory from time–frequency representations of the binaural signals. The system is systematically evaluated with respect to its input representation, recurrent architecture, and training procedure. In addition, a new synthetic dataset is introduced in which head-motion trajectories are generated from potentially overlapping minimum-jerk submovements subject to explicit motion constraints.
The experiments show that binaural-cue features with larger short-time Fourier transform configurations and increased recurrent depth improve the estimation performance. The final system achieves a median angular root-mean-square error of approximately 10◦ on recorded binaural data, representing a reduction of roughly 40% relative to the previous system on the same evaluation condition.
The synthetic training distribution is found to have a strong influence on the learned estimator, while the minimum-jerk dataset does not provide a clear advantage on the available recorded data. Furthermore, the ambiguity between opposite rotation directions can be resolved reliably on synthetic data when flip compensation is disabled. On recorded signals, however, ambiguity resolution is substantially less consistent and shows a pronounced listener dependence. These results indicate that reliable head-rotation estimation from binaural signals is feasible, while robust ambiguity resolution on measured data remains an open challenge.
