In this paper we study how sound quality is evaluated by different groups of assessors, with different levels of hearing loss. Formal listening tests using the Basic Audio Quality scale were designed using 22 headphones spanning a wide range of qualities and sound quality characteristics. The tests were performed with two formally selected listening panels with normal hearing (NH), and mild (N2) or moderate (N3) hearing loss characteristics. It is shown that not only do the two panels evaluate the sound quality consistently within each panel, but also that there are systematic changes in the manner in which hearing loss impacts the evaluation and ranking of the devices under study. Using this data we successfully train machine learning algorithms to predict the sound quality for the two assessor type panels. The prediction performance for each panel is NH: RMSE = 7.1 ± 3.0, PCC = 0.91 ± 0.13; HI: RMSE = 8.7 ± 2.4, PCC = 0.91 ± 0.12. Whilst it may not be practical to run listening tests with multiple panels of assessors, we demonstrate here that machine learning based models can be practically and cost effectively employed to predict the perception of multiple assessor groups rapidly and simultaneously.
(See document for exact affiliation information.)
https://aes2.org/publications/elibrary-page/?id=21087
(365KB)
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member Join the AES. If you need to check your member status, login to the Member Portal.
