Scene-aware compensation strategies for hearing aids in adverse conditions

The primary functionalities offered by hearing aids include speech enhancement (SE) and wide dynamic range compression (WDRC). SE aims to attenuate background noise and reverberation, while WDRC seeks to restore the loss of audibility by amplifying soft sounds without excessive amplification of loud sounds. Ideally, a combination of these building blocks can improve the listening experience in adverse acoustic conditions.
However, the two building blocks can have opposing effects. The SE might erroneously attenuate soft speech components, while the WDRC stage can amplify residual noise and artifacts from the SE stage.
Some of the opposing effects can potentially be reduced through either adaptive WDRC systems that are controlled by the speech presence or through simultaneously providing WDRC and SE in a joint configuration implemented as a deep neural network (DNN). However, many implementation and configuration options exist for both SE and WDRC.
This thesis explores the evaluation and comparison of hearing aid algorithms, primarily in terms of their similarity to an assumed ideal hearing aid that can provide both SE and WDRC without the opposing effects. The SE and WDRC systems were evaluated objectively through physical metrics and subjectively through listening experiments. First, a data-driven distance metric was proposed for the objective evaluation using physical metrics to evaluate the similarity of several WDRC systems. This data-driven metric consisted of three interpretable factors, where the first factor described the amount of compression applied to the target signal, the second factor described the compression and distortion of the background signal, and the third described the change in signal-to-noise ratio (SNR). Based on these factors, a smaller set of physical metrics was selected to evaluate the combined effects and the interactions of multiple configurations of both SE and WDRC.
The physical metrics were related to perception through two behavioral experiments. Here, a selected number of the considered systems were evaluated in terms of their relative preference and similarity. The similarity data was used to determine a set of perceptual attributes that represented how people with sensorineural hearing loss would judge the characteristics of the evaluated systems. The physical metrics were then related to the perceptual attributes, thus providing insights into how the physical metrics relate to perception.