Gedistribueerde algoritmen voor ruisonderdrukking in draadloze akoestische sensornetwerken

The increasing proliferation of audio-capable devices in modern environments has led to a growing interest in wireless acoustic sensor networks (WASNs). These networks leverage the wireless communication capabilities of spatially distributed devices to collaboratively solve audio signal processing tasks. Compared to traditional fixed microphone arrays, WASNs offer advantages such as flexible deployment, scalability, cost-effectiveness, and potentially improved performance, with applications ranging from environmental monitoring and smart homes to healthcare and industrial automation. Despite their promise, the practical use of WASNs requires overcoming significant challenges. Algorithms originally designed for single-device processing are not straightforward to deploy in distributed settings, where they must contend with limited communication bandwidth, constrained computational resources, time-varying network conditions, and potential asynchronicities between devices. A lack of efficient and robust algorithms that explicitly account for these challenges remains an important limitation to the widespread adoption of WASNs. This thesis focuses on distributed signal estimation, with a particular emphasis on noise reduction for speech enhancement, which is a task central to applications such as smart-home voice communication, cooperative hearing aids, and teleconferencing. Existing distributed solutions suffer from limitations in robustness, adaptability, and efficiency. The goal of this research work is to develop novel algorithms that aim at overcoming these limitations and thereby enhance the practical applicability of distributed signal estimation. The contributions of this thesis are threefold. First, the limitations of the topology-independent distributed adaptive node-specific signal estimation (TI-DANSE) algorithm in terms of long-term stability and slow convergence are addressed, leading to the development of a generalized eigenvalue decomposition (GEVD)-based version of the TI-DANSE algorithm, referred to as TI-GEVD-DANSE (for improved robustness), and of the TI-DANSE+ algorithm (for faster convergence). Second, new iterationless algorithms are proposed to provide statistically optimal estimators without iterative updates, enabling faster adaptation to changing acoustic conditions. These include the distributed multichannel Wiener filter (dMWF) for fully connected WASNs and the topology-independent dMWF (TI-dMWF) for ad-hoc networks under less restrictive inter-device connectivity assumptions. Third, the issue of sampling rate offsets (SROs) is tackled by integrating an SRO estimation and compensation mechanism into the DANSE framework, allowing robust performance in the presence of asynchronicities. In summary, this thesis contributes both extensions of existing approaches and entirely novel formulations that bring distributed signal estimation in WASNs closer to real-world deployment. Theoretical validity is established through formal analysis and proofs, while extensive numerical experiments confirm the effectiveness of the proposed solutions under diverse acoustic scenarios. Together, these contributions provide a foundation for scalable, flexible, and robust distributed speech enhancement systems.