Colloquium - Details
Bachelor-Presentation: Deep Learning-based Binaural Direction-of-Arrival Estimation Using Head Rotation Information
Johannes Holschbach
Monday, October 26, 2026
15:00 PM
IKS 4G | hybrid
Binaural Direction of Arrival (DoA) estimation is affected by ambiguities that arise because different source directions can produce similar interaural cues. Head rotation can reduce these ambiguities by providing additional observations of the source from different orientations.
This thesis investigates how head rotation influences binaural DoA estimation and whether the resulting error structure can be explained by the geometry of the Cone of Confusion (CoC).
Two approaches are compared: the model-based Movement-Invariant Beamforming (MIBF) method and a Neural Network (NN) estimator that combines binaural audio features with head-orientation information. Both methods are evaluated on generated binaural data and on real recordings. In addition to the overall angular localization error, the analysis uses the Cone of Confusion Distance (CoCD) and motion-dependent confusion points derived from the residual ambiguities of pure roll and yaw rotations.
For MIBF, the results show clear rotation-axis-dependent behavior. Yaw strongly reduces azimuth errors, while roll substantially reduces elevation errors on the real recordings. Its remaining errors also show a structure consistent with the predicted CoC-related ambiguities.
The NN achieves substantially lower localization errors on the generated data, but shows only weak dependence on the rotation axis. This advantage does not transfer to the real recordings, where localization errors increase strongly and performance varies considerably between participants.
Overall, the results show that providing head-rotation information to the NN does not by itself imply that the geometrically relevant information is exploited in an interpretable way. They also highlight the importance of evaluating both localization accuracy and the structure of localization errors under real recording conditions.
