Classification of Environmental Sounds for Future Hearing Aid Applications
Authors: Jürgen Tchorz, Simone Wollermann, Hendrik Husstedt
Abstract:
Different acoustic environments require different hearing aid settings to achieve best speech understanding and sound quality. Manual adjustment of hearing aid settings can be annoying. Thus, many hearing aids automatically classify the acoustic environment and switch between different programs accordingly. The classification approach presented in this study utilizes so-called amplitude modulation spectrogram (AMS) as features, which replicate aspects of sound analysis in the auditory pathway. The AMS patterns represent time intervals of 500 ms each. The classification of the acoustic environment based on these features is implemented with supervised machine learning using a deep neural network. The network is trained on features extracted from several hours of sound from different classes, namely speech, reverberant speech, speech in noise, music, and noise. For testing, a set of sounds taken from other recordings was processed and classified by the neural network. For comparison, these sounds were also automatically classified using hearing aids from five different brands. The results show comparable classification accuracy with amplitude modulation spectrograms and hearing aids, respectively. The time which is needed to classify a situation, however, is much shorter with the amplitude modulation spectrogram-based approach.