Hearing Aid Audio Processing Model Benchmarking with Binary Environmental Classification

Abstract

Current hearing aid audio processing models are trained to filter out noise and amplify specific sounds such as speech. However, real-world audio environments contain a multitude of sounds that users may want to hear that cannot exhaustively be defined within the model. Team ECHO’s project proposes a method for hearing aid audio processing models to account for these unknown sounds without explicit definition. The team developed an open source audio dataset containing unfiltered-environmental audio that is categorized by a set of 4 binary features (indoors, crowded, walking, speaking) which are applicable to any audio environment. The team collected audio data using over-the-ear microphones at different locations centered around the UMD College Park campus and in Washington DC. The team also constructed a benchmarking platform for researchers to compare the performance and efficiency of different models in different environments based around this dataset. The platform evaluates models on 6 different metrics (e.g., Noise Floor, Signal-to-Noise Ratio, Dynamic Range, and Crest Factor). Results from testing various audio processing models demonstrated significant differences in performance metrics across unique noise environments. The developed and tested benchmark serves as a groundwork for future audio processing research tailored specifically for hearing aid comfort and realism.