Full Name: Asaf Levhar Full Name: Michael Tatarjitzky Full name: Tamara Gucovschi

Full Name: Asaf Levhar Degree: Master of Science (M.Sc.) Advisor Name: Alon Kuperman Seminar Topic: Load-Side Control for a Battery-Integrated Dual-Winding Magnetic Energy Harves Seminar Abstract This seminar focuses on a smart way to "harvest" power from high-voltage power lines without any physical connection. Imagine a device that simply clamps onto a power line and uses its magnetic field to generate clean energy for remote sensors or equipment. We developed a new system that uses a dual-winding design to manage this energy more efficiently. The main challenge was to power a local device while simultaneously charging a backup battery and ensuring the voltage stays stable. By using advanced control strategies, we proved that this system can reliably provide power even when conditions change, making it a sustainable solution for monitoring electrical grids in remote areas. Full Name: Michael Tatarjitzky Degree Program: M.Sc. in Electrical Engineering Supervisor: Prof. Boaz Rafaely Seminar Title: Array-Agnostic Speech Enhancement Using Ambisonics Encoding and Dropout-Based Learning Seminar Abstract: Multichannel speech enhancement exploits spatial information to improve speech intelligibility and quality. However, most learning-based approaches are tailored to specific microphone array geometries and struggle to generalize when the array configuration changes. Existing array-agnostic methods typically rely on large datasets covering many array geometries, yet they may still fail to generalize to unseen layouts. In this seminar, we present AmbiDrop (Ambisonics with Dropouts), an Ambisonics-based framework that enables array-agnostic speech enhancement. The proposed method encodes arbitrary microphone array recordings into the spherical harmonics domain using Ambisonics Signal Matching (ASM). A deep neural network is trained on simulated Ambisonics data while incorporating channel-wise dropout to improve robustness to array-dependent encoding inaccuracies, eliminating the need for extensive multi-geometry training data. Experimental results demonstrate that while baseline methods and AmbiDrop perform similarly on arrays seen during training, baseline performance degrades significantly on unseen array configurations. In contrast, AmbiDrop consistently improves SI-SDR, PESQ, and STOI scores, highlighting its strong generalization capability and practical applicability to array-agnostic speech enhancement Full name: Tamara Gucovschi Supervisor: Prof. Joseph Tabrikian Degree: M.Sc. student in Electrical and Computer Engineering Seminar topic: Bayesian Performance Bounds Under Model Misspecification Seminar abstract: Performance bounds provide theoretical limits on the achievable accuracy of estimators under a given statistical model. However, in many practical estimation problems, the assumed statistical model differs from the true model, leading to model misspecification. This seminar focuses on Bayesian performance bounds under model misspecification. First, the Misspecified Cramér-Rao Bound (MCRB) and existing Misspecified Bayesian Cramér-Rao Bounds (MBCRBs) are reviewed, together with their main limitations. Then, a new pseudo-true parameter is derived using a transformation of the assumed probability density function. Based on this pseudo-true parameter, a revised MBCRB is proposed. The proposed bound is estimator-independent, accounts for prior statistical information, and coincides with the Bayesian Cramér-Rao Bound (BCRB) in the matched case. Finally, simulations for a scalar linear Gaussian model are presented to demonstrate the behavior and tightness of the proposed bound.