Project info
Project Title: Novel sparsity-aware adaptive algorithms for acoustic applications
Code: No. 112/2022
Program name and purpose: PCE - Supporting and promoting fundamental, multi-, inter- and trans-disciplinary and/or exploratory scientific research in Romania. The program is addressed to researchers with performances demonstrated by the quality and international recognition of scientific publications.
Funding Institution: Executive Unit for Higher Education, Research, Development and Innovation Funding (UEFISCDI)
Project code: PN-III-PCE-2021-0438
Project outcome
In many applications related to the acoustic devices, there is a need to estimate an acoustic impulse response, which results due to the acoustic coupling between loudspeaker and microphone. The practical solution is to use an adaptive filter working in a system identification scenario. In this project, we have developed new adaptive algorithms for acoustic applications, by following two main ideas. (i) First, the sparse nature of the acoustic impulse responses has been exploited. Most of the state-of-art algorithms perform well for very sparse impulse responses, but they usually fail when the number of non-zero coefficients increases (like in acoustic scenarios). The project has provided a novel framework to design sparsity-aware algorithms, by integrating an individual control approach based on specific optimization criteria (so that even the small coefficients could contribute to the overall convergence) together with the model uncertainties (related to the time-variant characteristic of the system). (ii) Second, a critical issue is related to the long length of the acoustic impulse response (hundreds or even thousands of coefficients). To this purpose, we have followed a novel approach, by exploiting tensor-based decomposition techniques, together with low-rank approximation methods. As a result, a high-dimension system identification problem is reformulated based on low-dimension problems (i.e., shorter filters) that are combined together. The gain is twofold, in terms of both performance and complexity.
Summarizing, the project has integrated several original ideas in order to overcome some of the main limitations of the current state-of-the-art solutions, by using: (i) the "individual control" approach based on rigorous optimization criteria, including specific constraints into the optimization problem (instead of the classical "proportionate" approach, which is developed in an ad-hoc manner); (ii) the time-variant characteristic of the acoustic impulse response based on a first-order Markov model (instead of the time-invariant assumption, usually involved in most of the developments); (iii) the non-parametric feature of the designed algorithms (which is critical in real-world applications); and (iv) the tensor-based decomposition approach, involving the Kronecker product together with low-rank approximation (which is a novelty in acoustic applications).