2022 Festive Hearing SIG Webinar
15th December 2022, 12-1pm (UK).
Registration for the webinar is via this link:
https://us06web.zoom.us/webinar/register/WN_KRnN5XZQQ52xkPgGAIvI7Q
After registering, you will receive a confirmation email containing information about joining the webinar.
Machine Learning Challenges to Improve Hearing Devices: The Clarity and Cadenza Projects
Trevor Cox and Becky Vos, University of Salford
Clarity and Cadenza are two EPSRC projects that are exploiting the latest in machine learning to create improved listening experiences for those with a hearing loss. Clarity is focussing on speech-in-noise whereas Cadenza is concerned with music. For both projects, the research methodology is to run a series of open competitions, within which researchers from around the world take part. Competitors are given speech or music to enhance. Competitors are challenged to improve and personalise the audio for listeners with hearing losses.
The competition methodology fosters a new research community devoted to making music and speech more accessible, as well as creating open-source tools and databases to facilitate future investigations. There will be three challenge rounds posing a variety of problems to the challenge entrants. Clarity is focussed on hearing devices, whereas Cadenza considers both hearing aids and consumer devices. While a hearing aid must manipulate sound with low latency and limited computing power, recorded music from consumer devices can be pre-processed with non-causal techniques using cloud computing. For music, we will also consider cases where separate tracks are available, for example via the mixing desk in a live venue, and also pre-mixed music which is only available in stereo. Having separate tracks means aggressive unmixing algorithms are not needed, but often only stereo is available.
We will need to develop better objective models for evaluating the speech intelligibility and the audio quality of music. These are needed to guide machine-learning optimisation and to score challenge entrants. For music, a focus group of individuals with diverse hearing will use descriptive elicitation and sensory profiling to develop scales for rating samples. These scales will then be used by a larger panel of listeners to rate the audio sent in by entrants to the challenges. For speech-in-noise, testing of word recognition is used in listening tests. These psychoacoustic experiments will allow us to develop better objective models for listener in-the-loop optimisation and so improve sound for aurally diverse populations.
Clarity is now running an ICASSP SP Grand Challenge on speech enhancement for hearing aids, and in Spring 2023 will launch a prediction challenge to improve speech intelligibility modelling.
Cadenza first challenge will launch in Spring 2023, with one track trying to improve the processing of music for listening in a car, and a second track on listening to recorded music via headphones.