Cadenza Project
Machine Learning Challenges to improve music for people with hearing loss
Improving music listening for those with hearing impairment
Cadenza is a 4.5-year project that aims to improve music listening for people with hearing impairments. It addresses the need for better music processing through machine learning challenges.
In the duration of this project, we ran five challenges covering issues in music audio quality and lyric intelligibility.
CAD1 - The first Cadenza Challenge
CAD1 was the first challenge presented in March 2023.
It included two listening scenarios: (A) Music over headphones, and (B) Music over car’s loudspeakers.
(A) The scenario considered a listener listening to music over headphones without wearing their hearing aid, so the headphones had to compensate for the hearing loss. The challenge was presented as a demixing/remixing task in which participants had to demix stereo tracks into vocal, drums, bass and other (``VDBO’’) stems, to allow these to be remixed in a way that improves audio quality. Unlike traditional demixing challenges, this task used HAAQI to evaluate the VDBO stems and the remix. Additionally, the remixed signal were evaluated by a listening panel using audio quality scales developed withing the project.
(B) This scenario was about listening to music over the car’s loudspeakers in the presence of car noise. Here, listeners are wearing fixed hearing aids. Participants’ algorithm had to process the music played by the car stereo in a way that increases its audio quality. For this, participants have access to: (i) The clean reference song. (ii) The audiogram of the listener. (iii) The car speed and noise metadata; participants did not have access to the temporal fine structure of the noise. The output signals were also evaluated using both the objective metric HAAQI and a listening panel.