Detecting the sounds that indicate a loss of autonomy

The coffee maker bubbles later and later. The microwave oven does not hum at a fixed time. The shower flows a little less often. These small, banal slips in a daily routine are sometimes the early signs of a gradual loss of autonomy in elderly people. This may be the time to consider a home care assistant or to adapt the accommodation. In order to help families immediately identify the onset of this stage, the Sonaide project is designing a new service based on the analysis of certain daily sounds.

It all began at the INRIA Nancy Centre, when Nicolas Turpault began his PhD thesis with the MultiSpeech research team. “Almost all the researchers in this group work on speech. With my supervisors Emmanuel Vincent and Romain Serizel, we decided to focus more on background noise. In theory, with artificial intelligence algorithms, we can recognise all sorts of sounds when they are isolated. But in real conditions, it didn’t work as well. For the algorithm to learn, it needs to be able to access a lot of data which has already been annotated by a human being. This is a meticulous and time-consuming task. My thesis research demonstrated that this learning is still possible with little amounts of data and especially little annotated data. With a hundred different vacuum cleaner sounds, we can recognise a vacuum cleaner in general.”

Invited by one of his co-supervisors, Nicolas also contributed to the organisation of DCASE, the major international ambient sound classification competition. “I oversaw the test to detect ambient sounds in a domestic environment. This put me in touch with scientists from around the world, in addition to major companies working on these subjects. We generally arrived at the conclusion that these works lead to few applications that can be used by a wide public. And what actually interested me, beyond the research, was to produce something that might be truly useful. A service that could have a real societal impact. Even before my thesis, when I graduated from engineering school, I was already planning to create a business one day.”

© Inria - Photo Agence Ludys - A. Le Ny

The next chapter unfolded in Rennes last October. The INRIA centre organised its hackAtech, an innovation marathon to initiate deep tech companies based on the Institute’s scientific results. Nicolas Turpault’s team won the jury’s Grand Prize. This provided him with the opportunity to join the INRIA Startup Studio and move the project forward. “I didn’t write a single line of code in the first months. I focused on exploring the market. There were several ideas out there for applications based on the technology. In the end, we opted for the detection of loss of autonomy.”

A new milestone was arrived at in April 2022 with the recruitment of a deep learning engineer. The engineer took over the development aspect. An initial experiment was carried out in a seniors’ residence in Vitré, Ille-et-Vilaine.

In addition to respecting private life, the artificial intelligence developed in this project requires little computing and thus little energy. “At the moment, companies are designing huge models which have an enormous environmental impact. My thesis, on the contrary, explored the possibility of creating small models capable of operating with very little data.” A first demonstrator is currently in development.

At the same time Nicolas Turpault is meeting with potential partners. One to act as Technical Director, the other as Sales Director. The company is likely to be created in late 2022 or early 2023. The search for investors to accelerate the project can then begin.