REal-time artiFicial INtelligence for hEaring aiDs – REFINED

About 466 million people worldwide suffer from hearing loss. Of these, 34 million live in the EU and 6 million in France. With the aging of the world's population, this number is expected to increase to 900 million by 2050, WHO says. Still according to the WHO, disabling hearing loss can lead to depression, loneliness and social isolation, as well as reduced levels of employment and income. The cost associated with hearing loss is currently estimated to €216 billion per year in Europe (Hear-it, 2019) due to reduced productivity, reduced quality of life, and health and societal costs.

Current estimates suggest that in France (Kervasdoué-Hartmann report 2016), only 30-35% people who would need a hearing aid actually use one. Current hearing aids provide a poor experience in noisy environments or for multiple sound streams, which is detrimental to social communication. In addition, patients must return to the hearing care professional sometimes more than ten times a year (at the beginning) to adjust settings.
To improve the efficiency of hearing aids, one solution is to equip them with filtering/separation algorithms to isolate the relevant streams. Machine learning has made the use of these algorithms credible in real life conditions, in often-complex scenarios. However, current artificial intelligence methods are too complex to be applied to these portable devices equipped with processors with low computing and memory capacities. Moreover, existing filtering/separation algorithms are generally not adapted to the particular characteristics of the patients' hearing loss.

The REFINED project is based on the upstream identification of auditory and extra-auditory spectro-temporal cues that correlate with the level of speech perception in people with auditory neuropathy spectrum disorders, constituting the 10% fringe of subjects who are more in need of speech filtering/separation than the primary function of conventional hearing aids (sound amplification). We study and develop efficient algorithms for auditory stream separation based on machine learning. We simplify them in order to implement them in real time while maintaining their performance and test their effectiveness on a cohort of carefully selected volunteer patients. This is done in an original approach in constant interaction between algorithms, embedded development and tests on patients. We develop experimental strategies based on the knowledge of sound perception to tune the algorithms in order to optimize speech recognition with respect to the limitation of information transfer and processing.

The consortium brings together partners from three very different backgrounds, who will pool their respective expertise to design a system that naturally lies at the border of their three worlds. The Institute of Hearing (Pasteur Institute) ensures the selection of subjects with well-defined audiological profiles on whom to test speech-processing systems, the LORIA (University of Nancy) brings skills in machine learning applied to speech enhancement and sound source separation and the CEA (project coordinator) brings its skills in artificial intelligence and design of performance constrained embedded systems.
To the best of our knowledge, REFINED will be the first initiative aiming at implementing an end-to-end adaptive AI-based solution for patients suffering from hearing loss, with a focus on the embedding the developed algorithms.

WHO: www.who.int/deafness/estimates/en/
Hear-it, 2019: www.hear-it.org/untreated-hearing-loss-eu-costs-more-whole-eu-budget
de Kervasdoué, J., & Hartmann, L. (2016). Impact Economique du Déficit Auditif en France et dans les Pays Développés. UNSAF