Name: Sheli Hendel Degree: Electrical and Computer Engineering Supervisor: Boaz Rafaely Full name: Shay Galperin Degree: M.Sc Student - Electro-Optics Supervisor: Prof. Yitzhak Yitzhaky Student Tom Dadon Supervisor Prof. Amir Geva

Name: Sheli Hendel Degree: Electrical and Computer Engineering Supervisor: Boaz Rafaely Subject: Analysis of the trade-off in spatial and speech qualities of masking-based speech enhancement for spatial audio Abstract: Speech enhancement for spatial audio requires a delicate balance between noise suppression and the faithful preservation of spatial cues. While masking-based approaches are highly effective for single-channel enhancement, their performance in multi-channel spatial contexts is less understood. This paper presents a systematic analysis of the trade-offs between speech quality and spatial fidelity across three distinct processing domains: Time-Frequency Microphone-channels (TFM), Ambisonics (TFA), and Beamforming (TFB). Using a spherical microphone array framework and Monte Carlo simulations, we evaluate these methods under varying acoustic conditions, considering both direct and reverberant target signals. Our results reveal a fundamental trade-off: TFB processing yields superior noise suppression and objective speech quality (PESQ, SI-SDR), but significantly degrades spatial realism. In contrast, TFA provides the most robust preservation of interaural cues (ITD, ILD) and environmental attributes (DRR), making it better suited for immersive applications. We further demonstrate that defining the desired signal as the reverberant speech significantly improves the preservation of the acoustic scene compared to a direct-path target. Full name: Shay Galperin Degree: M.Sc Student - Electro-Optics Supervisor: Prof. Yitzhak Yitzhaky Subject of Research: Creating Streetscape Images with Controllable Perceptual Quality Parameters Using Generative AI Seminar summary: The research focuses on conditioning diffusion-based generative models on quantified streetscape perceptual attributes, moving beyond purely semantic or stylistic control. A novel, purpose-built dataset of street-level images was constructed and annotated with continuous urban metrics, including sky visibility percentage, building height–to–street width ratios, and structured streetscape features. These attributes are derived through a combination of manual geometric annotation and automated vision pipelines, producing interpretable, physically grounded supervision signals aligned with urban morphology. Using this dataset, diffusion models are fine-tuned with custom token embeddings that encode perceptual ranges as controllable conditioning variables. The study evaluates how token design, optimizer choice, and training regimes affect attribute fidelity and disentanglement at inference time. The results demonstrate that key perceptual urban variables can be learned as controllable latent factors, enabling generative models to function not only as image synthesis tools, but also as quantitative instruments for urban analysis and design exploration. Student Tom Dadon Supervisor Prof. Amir Geva Degree M.Sc. in Electrical and Computer Engineering Thesis Title Advancing Blood Pressure Pulse Waveform Decomposition: Model and Algorithm Abstract Cuffless blood pressure estimation has plateaued at 10–14 mmHg mean absolute error despite two decades of research, largely because most approaches treat the translation from peripheral waveform features to blood pressure as a black-box problem without addressing what hemodynamic information the arterial pulse actually encodes. This work advances the field by proposing a principled, physics-grounded decomposition of arterial blood pressure waveforms into their forward-propagating and reflected components, a prerequisite for more robust cardiovascular assessment. We introduce the Aorta Reflection Model (ARM), a simplified six-parameter representation of wave reflections from two anatomically motivated sites (renal and iliac bifurcations), and the Shift-Based Decomposition (SBD) algorithm, a template-free iterative method that simultaneously recovers the forward wave and quantifies reflection contributions without requiring a pre-specified forward wave shape. Physiological guardrails ensure that solutions remain hemodynamically plausible. Validation employs a novel three-level converging evidence framework — reconstruction fidelity, blood pressure estimation utility, and physiological corroboration through independent timing measures. The model and algorithm were validated over a set of 1.2 million heartbeats from 1128 subjects from ICU and surgical procedures in the PulseDB database (ill patients), along with independently collected 0.5 million beats from 205 healthy subjects. This work reframes cuffless blood pressure estimation as a two-stage problem: first understand what the waveform encodes through principled decomposition, then leverage that understanding for estimation.