Abstract
Objective A hearing aid’s noise reduction algorithm cannot infer to which speaker the user intends to listen to. Auditory attention decoding (AAD) algorithms allow to infer this information from neural signals, which leads to the concept of neuro-steered hearing aids. We aim to evaluate and demonstrate the feasibility of AAD-supported speech enhancement pipelines in challenging noisy conditions without access to clean speech signals.
Methods We evaluated a linear versus a deep neural network (DNN) based speaker separation pipeline, with same-gender speaker mixtures for 3 different speaker positions and 3 different noise conditions.
Results AAD results based on the linear approach were found to be at least on par and sometimes even better than pure DNN-based approaches in terms of AAD accuracy in all tested conditions. However, when extending the DNN with a linear data-driven beamformer, a performance improvement over the purely linear approach was obtained in the most challenging scenarios. The use of multiple microphones was also found to improve speaker separation and AAD performance over single-microphone systems.
Conclusion Our study shows that neuro-steered speech enhancement, combining the best of both worlds (linear and DNN), results in robust performance.
Significance Recent proof-of-concept studies in this context each focus on a different method in a different experimental setting, which makes it hard to compare them. Furthermore, their idealized experimental conditions only give a rather premature evidence on the viability of the AAD paradigm in a hearing aid context. This work provides a systematic proof-of-concept of neuro-steered speech enhancement in challenging conditions.
Footnotes
The work is funded by KU Leuven Special Research Fund C14/16/057, FWO project nr G0A4918N and the ERC (637424 and 802895) under the European Union’s Horizon 2020 research and innovation programme. Jeroen Zegers is supported by an SB PhD scholarship 1S66217N of FWO. The scientific responsibility is assumed by its authors.