Speaker
Description
Speech understanding in noise remains the primary challenge for individuals with hearing loss and a central target for hearing aid technology. Over the past forty years, advances in directional microphones and statistical noise cancellation have delivered substantial improvements in signal to noise ratio (SNR). More recently, deep neural network (DNN)–based speech enhancement has enabled a fundamental shift in algorithmic capabilities, offering robust performance even in complex and dynamic acoustic environments.Despite these technological advances, outcome measures used to assess hearing aid performance have changed little. Metrics such as speech intelligibility, SNR improvement, and user preference remain dominant. In addition, test setups often rely on static noise sources, limited spatial complexity, and favorable SNRs. While well suited to demonstrate the benefits of classical signal processing approaches, simplified test setups tend to overestimate algorithmic benefit and fail to reflect the benefit in real world listening.In this talk, I will demonstrate how the estimated SNR improvement depends on the spatial complexity, noise type, and input SNR of the background noise. Under simplified conditions, classical noise cancellation approaches can perform comparably to or better than DNN-based speech enhancement. However, in more complex environments, DNN based approaches provide superior benefit. The findings demonstrate that simplified test paradigms obscure the conditions under which DNN-based algorithms outperform classical methods. To keep pace with technological progress, the field is moving toward more ecologically valid test methods, advanced modeling approaches, and outcome measures that extend beyond SNR and speech intelligibility and toward quality of life relevant measures.