Who has never been bothered by noise in a refectory, swimming pool or hall? There are many overly noisy rooms. The only solution for lessening user discomfort is to carry out acoustic rehabilitation. In practice, the acoustician conducts a study to propose new wall, floor or ceiling coverings capable of efficiently absorbing sound.
To select these surfaces and area to cover, along with the necessary acoustic performance of the absorbent materials according to room usage, a reliable acoustic diagnosis of the existing structures must be carried out.
"The acoustician conducts a measurement campaign in the room using calibrated sound sources and sound level meters, explain Antoine Deleforge, initiator of the Acoust.IA project at Inria, and Cédric Foy, UMRAE/Cerema researcher associated with the project. He/she then uses digital acoustic simulation tools requiring the room's 3D geometry to be modelled and its input parameters to be defined from these measurements".
A few sound recordings to do everything, an ambitious goal
The difficulty lies in the fact that it takes days of work to adjust the input parameters so that the modelled sound field best matches the measurements made in-situ and then to virtually test different rehabilitation scenarios and select the one that seems most optimal. The bill for diagnosis alone can quickly reach the thousands of Euros.
The ambition of the Acoust.IA exploratory action, conducted by Inria (French national institute for research in computer science and control) and UMRAE (Joint research unit in environmental acoustics), bringing together researchers from Cerema and the Gustave Eiffel University (formerly Ifsttar), is to automate this process by means of artificial intelligence and machine learning:
"We want to be able to enter a room with an acoustic recorder and, why not, eventually a smartphone and application, generate a series of calibrated sounds or simple signals such as claps, and to automatically obtain the 3D geometry of the premises, along with the absorption profiles of all coatings present"
An interdisciplinary approach combining acoustics and computer science
Inria and UMRAE accepted this challenge by pooling their complementary acoustics and computer science skills: the Inria Multispeech team specialises in automatic learning for sound processing and the spoken word, while UMRAE specialises in particular in acoustic propagation modelling and possesses extensive experience in building acoustics.
Historically, computer science and acoustics have been highly disjointed research fields, between which collaborations have been rare. Through the Acoust.IA exploratory action, common methodology bases will be developed to address this problem.
Moreover, for the first time a scientific article published in 2014 proposed a method for estimating the geometry of a room from wall reflections arising from the propagation of well-calibrated sounds.
Knowledge of room geometry, however, is essential in setting up the diagnosis as it provides a better understanding of how the sound field propagates and which area should primarily be covered with absorbent materials. Unfortunately, most of the time, the acoustician does not have plans when it comes to rehabilitation projects or they have not been updated after possible work.
The 2014 method, however, only works under perfectly controlled and calibrated conditions, whereas Acoust.IA aims for a solution applicable everywhere and using simple equipment capable of estimating:
-
the geometry of the most complex rooms
-
the absorption profiles of wall coverings, even if several are used (concrete, tiles, wood, glass, fabric, etc.).
To achieve this, the researchers will develop an automatic tool resulting from supervised learning, but also from taking into account the physical models of sound propagation allowing the sound field to be estimated, to within a millisecond, from the model sound propagation paths resulting from reflections on wall coverings. "UMRAE has hundreds of sound recordings in premises of well-known dimensions. We will "teach" these rooms to our tool in order to calibrate it little by little while integrating the knowledge provided by the physical models".
Neural networks to determine surface absorption profiles
The estimation of wall covering absorption profiles is the focal point of the project. How can we determine the absorption profiles of the room's surfaces? Admittedly, there are reliable models for estimating the sound field from the propagation paths resulting from reflections off the walls, but there are no inverse models capable of returning to the absorption profiles of wall coverings from the sole knowledge of the sound field, due to the mathematical difficulties which then arise.
To carry out an acoustic diagnosis of an existing room with a view to its rehabilitation, the acoustician must be able to estimate the acoustic performance, in terms of absorption, of the wall materials used. To date, the direct on-site measurement of this sound absorption remains too complex to implement. The acoustician then measures the sound field (a) from which he/she seeks to deduce this acoustic absorption provided by the wall materials.
For this, he/she uses numerical acoustic forecasting tools. Unfortunately, to date, the only existing tools solve the opposite problem, i.e. estimating the sound field from a known wall absorption. The only way for the acoustician to use these tools is then to carry out several successive simulations by gradually modifying the wall absorption in order to match the measurement as closely as possible. The current approach is thus long and fastidious.
The Acoust.IA project aims to overcome this difficulty by offering a tool capable of directly estimating wall absorption from the measurement. It is based on the idea that the "wall absorption" information is necessarily "nested" to the extent of the sound propagation paths based on the multiple reflections of sound with the walls as shown schematically above (b) The main objective of the Acoust.IA project is to use artificial intelligence to find this "nested" information!
Automated learning, still underused in the field of acoustics, has great potential for solving non-linear inverse problems such as this one. Indeed, using a large set of so-called training input and output data, these methods are used to estimate a non-linear function connecting the two, for example with the help of a deep neural network.
With the ACoust.IA project, for the first time in acoustics, researchers will exploit specific neural networks called "variational autoencoders". They are able, using probabilistic models, to calculate a result from input data and to find input data from a result.
The Acoust.IA exploratory action began on 1st October, and a thesis co-supervised by Inria, UMRAE and the University of Strasbourg (ICube) was launched, with the objective of developing a system to deal, more quickly, at a lower cost and with better results, with the noise pollution which today weighs on a large part of the population.