Listening test are usually time-consuming and costly. Maintaining a panel of highly trained assessors requires significant resources. This study introduces a neural network model, the Virtual Listener Panel, trained using Convolutional Neural Network to overcome these challenges. The focus in this study has been to predict expert ratings for audio characterisation using deep learning. This paper describes the architecture of the proposed model and the data collection approach that was established to achieve high prediction accuracy. The performance of the Virtual Listener Panel is compared with data from an independent listening test where seven different earbuds were characterised by 20 trained expert assessors. Five audio characterisation attributes from the timbre domain are validated: bass strength, bass depth, midrange strength, treble strength and brilliance. The result of this study demonstrates Pearson correlation coefficients between 0.80 and 0.97. The Virtual Listener Panel offers a scalable solution for audio characterisation, reducing the need for extensive listening tests and providing valuable insights for product development.
(See document for exact affiliation information.)
https://aes2.org/publications/elibrary-page/?id=23014
(679KB)
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.
