Psychoacoustic Annoyance and the EEG – A Deep Learning Approach
* Presenting author
Abstract:
Jury testing and the application of psychoacoustic metrics to sound pressure data are two established methods for the assessment of sound quality. So far, little is known to what extend the results of these methods correlate with biosignals, especially the electrical activity of the brain.A growing number of manufacturers offers wearable biosensors at a relatively low cost. In this contribution, it will be discussed whether the psychoacoustic annoyance of a particular sound (Zwicker’s PA index) can be predicted based on measurements with a semi-professional EEG headset.An EEG study with 11 participants was conducted. 100 machine sounds and nature sounds of 4 s length were presented to each test person. An 8-channel EEG headset was used to record the EEG data synchronous with the acoustic stimulation.A deep learning approach for EEG classification based on a convolutional neural network was implemented to learn the class affiliations PA < mean(PA) and PA ≥ mean(PA) based on spectral entropy features. Despite the relatively poor validation accuracy of 58.8 %, this method proved to be sensitive to changes in the PA, which encourages further research with an optimized study design.