Contribution
Effects of Errors in Direction-of-Arrival Estimates and Head Movement Measurements on Directional Multi-Talker Speech Enhancement in Hearing Aids
* Presenting author
Abstract:
Speech enhancement is a critical aspect of hearing aid technology, particularly in challenging multi-talker environments. One approach utilized in so-called behaviour-controlled hearing aids is to leverage eye-gaze as an indicator of spatial attention, combined with estimates of the direction of arrival (DOA) of sound sources present in the acoustic scene, to dynamically steer directional filters towards relevant sound sources. Previous work in our group has demonstrated the potential of multi-beam Minimum Variance Distortionless Response (MVDR) beamforming to enhance multiple target speech sources and reduce noise in dynamic scenarios, including the listener's head movements. However, this work relied on oracle knowledge of DOA and head movements, raising questions about the algorithm’s effectiveness in real-world conditions.The present work extends the evaluation by incorporating estimation errors in terms of variance and biases to investigate how inaccuracies in DOA estimation and measured head movements affect the signal enhancement performance. By systematically replacing oracle data with error-prone data in simulations of realistic acoustic scenarios, the robustness of the employed multi-beam MVDR algorithm in real-world conditions can be assessed. Additionally, insights will be provided into the critical thresholds for acceptable error sizes needed for effective directional filtering, which is essential for the reliability of behaviour-controlled hearing aid algorithms in practice.