Experimental acoustic sensor networks are currently tested in large cities, and appear more and more as a useful tool to enrich modeled road traffic noise maps through data assimilation techniques. One challenge is to be able to isolate from the measured sound mixtures acoustic quantities of interest such as the sound level of road traffic. This task is anything but trivial because of the multiple sound sources that overlap within urban sound mixtures. In this paper, the Non-negative Matrix Factorization (NMF) framework is developed to estimate road traffic noise levels within urban sound scenes. To evaluate the performances of the proposed approach, a synthetic corpus of sound scenes is designed, to cover most common soundscape settings, and whom realism is validated through a perceptual test. The simulated scenes reproduce then the sensor network outputs, in which the actual occurrence and sound level of each source are known. Several variants of NMF are tested. The proposed approach, named threshold initialized NMF, appears to be the most reliable approach, allowing road traffic noise level estimation with average errors of less than 1.3 dB over the tested corpus of sound scenes.

Sound source detection and recognition using acoustic sensors are increasingly used to monitor and analyze the urban environment as they enhance soundscape characterization and facilitate the comparison between simulated and measured noise maps using methods such as Artificial Neural Networks or Non-negative Matrix Factorization. However, the community lacks corpuses of sound scenes whose acoustic properties of each source present within the scene are precisely known. In this study, a set of 40 sound scenes typical of urban sound mixtures is created in three steps: (i) real sound scenes are listened and annotated in terms of events type, (ii) artificial sound scenes are created based on the concatenation of recorded individual sounds, whose intensity and duration are controlled to build scenes that are as close as possible to the real ones, (iii) a test is carried out to validate the level of their perceptual realism of those crafted scenes. The interest of using such corpus is then demonstrated using an important task in urban environment description: the estimation of the traffic level in urban acoustic scenes.

In many countries, the acoustic impact of wind farms is often constrained by a curtailment plan to limit their noise, which spreads in their surroundings. To update the plan, on/off cycle measurements are performed to determine the ambient noise (wind turbines in operation) and residual noise (wind turbines shut down), but these shutdown operations are limited in time, which reduces the representativeness of the estimated in situ emergence. Consequently, a machine learning technique, called nonnegative matrix factorization (NMF), is proposed to estimate the sound emergence of wind turbines continuously, i.e., without stopping the machines. In the first step, the application of NMF on a corpus of various simulated scenes allows the determination of the optimal setting of the method to better estimate the sound emergence. The results show the proper adaptation of the method with regard to the influence of the propagation distance and atmospheric conditions. This method also proves to be efficient in cases in which the real emergence is less than 5 dB(A) with a mean error lower than 2 dB(A). The first comparison with in situ measurements validates these performances and allows the consideration of the application of this method to optimize wind farm operations.