SENTHIPoli Research Results

What We've Published So Far

Scientific articles resulting from the SENTHIPoli research project, published in journals and presented at specialized conferences.

18publications
4journals
10conferences

18 publications

Real-Time Detection and Masking of Romanian Offensive Words in Audio Streams Using ASR

Vicentiu-Lucian Milcu, Costin-Emanuel Vasile

ISETC26 - the 17th International Symposium on Electronics and Telecommunications · 2026

ConferenceTo be published

New Programming Approaches for Hardware Accelerators

Abstract

Automatic moderation of spoken content requires not only detecting offensive language, but also locating the corresponding portion of the audio signal so that it can be removed or masked. This paper presents a real-time system for detecting and masking predefined offensive words in Romanian audio streams using automatic speech recognition (ASR). The input audio is processed in partially overlapping chunks and transcribed using a Romanian FastConformer-based ASR model. Word-level temporal positions are estimated from CTC output frames and refined using local signal-energy variations. To increase robustness against transcription and pronunciation variations, recognized words are compared against an external blocklist using approximate string matching, after which the detected audio intervals are replaced by a masking tone. Experiments on controlled and natural Romanian speech show one-pass F1 scores of 80.8% and 75.0% on two natural-speech recordings, while the average ASR inference time remains below 110 ms.

Improving Graph-Based Sleep Staging with Shared Waveform Node Representations

Simona Juvină, Ana Neacșu Nicolae

Machine Learning and Knowledge Extraction · 2026

JournalQ1IF 8.4

Methods for Training Secure Artificial Intelligence Models

Abstract

The study investigates the impact of node representation on the performance of graph models for sleep stage classification. The specific frontends of five architectures are replaced with the same one-dimensional convolutional network, trained directly on polysomnographic signals. Three training strategies are analyzed on five datasets under comparable experimental conditions. The results show improvements for most model–dataset combinations, with the most consistent gains observed for the two-stage training strategy, particularly for the N1 and REM stages. The learned representations retain their usefulness even when transferred between datasets.

Towards Zero-Shot Attribution of Synthetic Speech via Audio-Text Contrastive Retrieval

Cristian-Teodor Neamtu, Serban Mihalache, Stefan Smeu, Dan Oneata, Horia Cucu, Dragos Burileanu

IEEE Spoken Language Technology 2026 · 2026

Conference

Speaker Identification in the Automotive Environment for Speech-Based Control with Authentication

Abstract

In this study, we didn’t stop at a simple „real or fake” verdict; instead, we wanted to determine directly which system generated the audio file. To avoid the pitfalls of traditional classification, we reframed the problem as a text-to-audio search task. Essentially, we describe each generator in simple terms and align the audio samples with the texts in a shared embedding space. This way, adding a new model no longer requires retraining. On MLAAD v9, we achieved an MRR of 58.4% on systems that were completely unseen during training. Even when the exact model is not identified, our method correctly recognizes the architecture, vocoder, or basic acoustic characteristics.

On the Potential of Forward-Looking Automotive SAR for Target Separation in Elevation

Radu Otopeleanu, Alexandru Gîrdianu, Mădălina Ciucă, Andrei Anghel, Remus Cacoveanu, Filip-Alexandru Roșu

IEEE Radio & Wireless Week (RWW2027) · 2026

ConferenceTo be published

Classification and Focusing Methods for Automotive Radar Images

Abstract

Elevation resolution remains a weak point of automotive radars, although it is essential for distinguishing overhead structures from vehicles ahead and avoiding unnecessary braking. We propose synthesizing a forward-looking elevation aperture from the vehicle’s forward motion. We formulate a model of range variation and analytically derive the Rayleigh resolution limit in elevation. Simulations show that an aperture of ~10 cm separates two reflectors a few degrees apart. Rail-mounted measurements using an NXP FMCW radar and back-projection focusing support the feasibility of the approach. Monte Carlo tests indicate that automotive SAR in elevation offers a viable, low-cost complement to MIMO in elevation.

S2X-Net: Stereo-to-Multichannel Neural Upmixing Network

Marian Negru, Bogdan Moroșanu, Ana Neacșu Nicolae

IEEE Transactions on Audio, Speech and Language Processing · 2026

JournalQ1IF 5.2

Methods for Training Secure Artificial Intelligence Models

Abstract

This paper presents an automatic upmixing algorithm based on neural networks, capable of converting any stereo audio signal—such as those from movies, music, or animations—into a 5.1 or 7.1 multichannel format. The proposed method combines source separation models with principal and ambient component extraction, achieving results close to those obtained manually, according to subjective MUSHRA tests. Furthermore, the mathematical formulation used enables an almost lossless conversion back to stereo, as evaluated by SDR, SI-SDR, and MSE.

Anti-tamper Conductive Mesh for SiP Devices

D. C. Vasile, P. M. Svasta, R. C. Maxim

SIITME 2026 · 2026

ConferenceTo be published

Physical Security of Embedded Modules

Abstract

This paper proposes a protection structure against physical intrusions, applicable to SiP-type circuits. The structure was designed and simulated in ANSYS HFSS; the results demonstrated excellent performance in detecting physical intrusions across the entire operating temperature range.

Improved Range Separability and Maximum Velocity for Automotive FMCW Radar via Frequency-Stepped Chirp

Alexandru Gîrdianu, Filip-Alexandru Roșu, Andrei Anghel

European Microwave Week 2026 (EuRAD 2026) · 2026

ConferenceTo be published

Classification and Focusing Methods for Automotive Radar Images

Abstract

The performance of automotive radars is limited by memory and bandwidth, preventing simultaneous improvements in range resolution and maximum unambiguous speed. This paper investigates the Frequency-Stepped Chirp (FSC) signal for reallocating the parameters of automotive FMCW radars. By reducing the fast-time sampling and reallocating radar cube resources to the slow time, the maximum unambiguous speed doubles, while maintaining the cube size and maximum range. The synthesized bandwidth improves range separability. Target motion introduces a phase term that causes Doppler-dependent energy dispersion at high speeds. Simulations and real-world measurements confirm the benefits of FSC under fixed hardware constraints and identify conditions leading to significant degradation.

CineSubNet: A Multimodal Architecture for Text-Driven Cinematic Audio Source Separation

Marian Negru, Ana Neacsu Nicolae

34th European Signal Processing Conference · 2026

ConferenceTo be published

Methods for Training Secure Artificial Intelligence Models

Abstract

This paper proposes CineSubNet, a multimodal architecture based on textual information for CASS. Subtitles are automatically generated using an automatic speech recognition (ASR) system and integrated without requiring explicit alignment with the audio signal, thereby facilitating a practical and scalable approach. Evaluations conducted on all variants of the Divide and Remaster (DnR) benchmark show that CineSubNet consistently outperforms existing methods in the literature across all analyzed metrics. The results highlight the potential of textual information for the development of CASS toward more realistic and multimodal cinematic scenarios.

Towards Universal, Transferable, and Targeted Acoustic Attacks on ASR Systems

Emanuel Pantelimon, Vlad Vasilescu, Ana Neacsu Nicolae

34th European Signal Processing Conference · 2026

ConferenceTo be published

Methods for Training Secure Artificial Intelligence Models

Abstract

This paper proposes MuteOut, a method for generating universal adversarial perturbations that are applicable to multiple speech recordings and transferable across automatic speech recognition (ASR) models. The study analyzes the mechanisms through which common architectural features facilitate the transferability of these attacks, highlighting security vulnerabilities in ASR systems used in black-box mode.

Unfolding for Audio Quality Improvement in the Presence of Noise and Reverberation

Marian Negru, Ana Neacsu Nicolae, Corneliu Burileanu, Jean-Christophe Pesquet

34th European Signal Processing Conference · 2026

ConferenceTo be published

Methods for Training Secure Artificial Intelligence Models

Abstract

This paper proposes a unified algorithm that simultaneously addresses the problems of signal denoising and de-reverberation. The method combines physical models—by integrating additional information such as room impulse responses and noise standard deviations—with a data-driven machine learning approach to improve performance. Evaluations conducted on datasets containing both speech and music demonstrate the method’s versatility and its ability to be effectively applied to various types of audio signals.

RADAR-MAE: Masked Autoencoders for Instance Segmentation In Radar Point Clouds

Elena Beatrice Anghel, Nicolae Catalin Ristea, Andrei Anghel, Filip-Alexandru Roșu

IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2026) · 2026

ConferenceTo be published

Classification and Focusing Methods for Automotive Radar Images

Abstract

Understanding the scene is essential for autonomous vehicles, enabling motion prediction, collision avoidance, and trajectory planning. Automotive radars, which are robust against adverse weather and lighting conditions and widely integrated into modern vehicles, complement cameras and LiDAR. However, radar point clouds are sparse, noisy, and have low resolution. We propose RADAR-MAE, a self-supervised framework that combines masked autoencoder (MAE) pretraining with the Single-stride Sparse Transformer (SST) for radar point cloud segmentation. The autoencoder learns dense and discriminative representations by reconstructing masked tokens. Experiments on RadarScenes show that the method outperforms fully supervised approaches and the current state of the art in both segmentation quality and mAP.

A Cross-Session Benchmark for Device-Free Presence Detection Using Wi-Fi Channel State Information

Andrei Țaga, Răzvan Crăciunescu, Răzvan Mihai, Mihnea Udrea, Constantin Pintilie

2026 IEEE International Black Seas Conference on Communications and Networking (BlackSeaCom) · 2026

Conference

Opportunistic Radar for the Automotive Industry

Abstract

Wi-Fi Channel State Information (CSI) enables device-free presence sensing, but models trained in one session often degrade when the link geometry or room changes. We evaluate binary presence detection with a strict leave-one-session-out (LOSO) protocol over six sessions designed to induce controlled shifts (receiver displacement/rotation, height change, corner placement, non-line-of-sight (NLoS) obstruction, and cross-room transfer). Each sample is a beacon-window CSI amplitude map (52 subcarriers × 8 beacons) processed with per-window normalization. Using identical windows, we benchmark a lightweight convolutional neural network (CNN) against a radial basis function (RBF) kernel support vector machine (SVM) across three representations: RAW amplitude, RAW plus temporal differences, and compact two-dimensional fast Fourier transform images. Beyond mean accuracy, we report false positives and false negatives to expose failure modes under shift. Results show that cross-room transfer is consistently the hardest condition and that errors are dominated by missed detections, highlighting the need for lightweight calibration or domain-robust training for deployment.

Benchmarking Framework for Multimodal Large Language Models in Image Understanding and Processing

Costin-Emanuel Vasile, Sophia-Anca Minișan, Călin Bîră

MDPI Journal of Imaging

JournalQ2IF 3.8To be published

New Programming Approaches for Hardware Accelerators

Abstract

This work introduces a configurable benchmarking framework for evaluating MLLMs in image understanding and enhancement. The framework supports quality assessment, no-reference enhancement, and reference-guided enhancement, while recording both image-quality metrics and agent-specific characteristics such as execution time, token usage, model requests, and tool interactions. We further propose an MLLM-driven image enhancement approach in which external image analysis and image processing operations are exposed as callable tools, allowing the model to iteratively select suitable operations and their parameters. The framework is demonstrated using a locally deployed general-purpose MLLM and blur and contrast distortions from KADID-10k. The results show that the effectiveness of enhancement depends heavily on the type of degradation, tool availability, and reference information, while unrestricted access to tools does not necessarily improve performance. The proposed framework provides a reproducible basis for studying and comparing MLLM-based image understanding and processing systems.

Hardware Acceleration for Physical AI: Shared Operations and Design Requirements

Răzvan Stancu, Alina-Luiza Petcu, Costin-Cosmin Dumitru, Han-David Zhao, Georgian Nicolae

Electronics

Journal

Extending the RISC-V instruction set for domain-specific applications

Abstract

Embedded physical AI combines sensing, learned inference and physical action under resource and timing constraints. This focused critical review examines how selected multimedia, AI/ML, radar/LiDAR and actuator-control workloads translate into hardware accelerator requirements. Its unit of analysis is a configured workload with explicit dimensions, numerical semantics, data movement, state, dependencies and execution phases. Pairwise records assess a candidate capability as conditionally compatible, requiring adaptation, incompatible under specified constraints, or insufficiently documented. The synthesis exposes FFT format and capacity mismatches, mutable filtering state and control-solver phases that change acceleration priority. Robotic examples connect classical transforms to learned action representations, distinguish action-generation throughput from observation-response latency, and document existing reuse across optimization algorithms. Physical motor-control evidence adds synchronized acquisition, persistent state and candidate selection. Shared mathematical operations, hardware-design reuse and use of one physical resource are evaluated separately. Sensor age, policy invocation, action production and actuator updates remain distinct timing requirements, with RISC-V as an implementation perspective. The resulting framework links proposed capabilities to source evidence and explicit acceptance conditions. Its compatibility judgments guide architectural investigation; demonstrating a benefit from shared hardware still requires matched numerical behavior, task quality and concurrent timing evidence.

MicroSlicer: Efficient Inference-Time Slicing for Small-Target Detection in Thermal UAV Search-and-Rescue Imagery

Andrei-Gabriel Manta, Costin-Emanuel Vasile, Călin Bîră

MDPI Journal of Imaging

JournalQ2IF 3.8To be published

New Programming Approaches for Hardware Accelerators

Abstract

Thermal drones help search for missing people in the wilderness, but flying high enough to cover wide areas reduces each person to just a few pixels, causing standard detectors to miss them. We introduce MicroSlicer, a lightweight method that improves small-target detection by slicing each frame into overlapping tiles at inference time, with no additional training required. Tested on three detector architectures (YOLO26-nano, YOLO26-small, RT-DETR-Large) using a thermal search-and-rescue dataset, MicroSlicer detects more small targets than standard full-frame inference and matches or outperforms an existing slicing method, SAHI, while running just as fast or faster. It performs particularly well on the transformer-based detector, where SAHI’s accuracy and speed both drop significantly at low confidence thresholds, whereas MicroSlicer remains stable. We also demonstrate that training directly on cropped patches is less effective than simply combining full-image training with MicroSlicer, resulting in up to a 24.8 percentage point difference in accuracy.

Presence detection of persons in a room using 5G ISAC non-ML techniques

Theodor Fratu-Halunga, Razvan Mihai, Dragos Nicolae Vizireanu, Razvan Craciunescu

2026 IEEE International Black Seas Conference on Communications and Networking (BlackSeaCom) · 2026

Conference

Opportunistic Radar for the Automotive Industry

Abstract

This paper proposes the implementation of a non machine learning based presence detector for people indoors, to be used on hardware platforms with limited compute resources, utilizing the Neyman-Pearson decision criterion in a bistatic SISO ISAC scenario. The decision criterion was calibrated and verified using a proof of concept testbed comprised of a OAIBOX gNB + 5GC combo and 5G enabled mobile device operating in the n78 band, by recording the Channel State Information extracted from uplink Sounding Reference Signals during trials conducted in Orange’s 5G Lab located in UNSTPB’s CAMPUS Research Institute. The measured data was split into frames and separated in two distinct ”presence” and ”no presence” datasets and, after analyzing their statistical distribution, a final detection rate of 99.69% was obtained with a false alarm probability of 1%.

Private Check Waveform Assisted False Target Rejection for OFDM ISAC

Laura Fricosu, Alexandru Vulpe, Astik Samal, Octavian Fratu

IEEE CloudNet 2026 · 2026

Conference

Radar and Sensing

Abstract

Integrated sensing and communication is increasingly used in 6G networks, but replay and spoofing attacks can create false targets in range-Doppler maps. Conventional CFAR detection finds strong peaks, but it cannot verify whether a detected target is real or attacker-generated. This paper proposes a lightweight trust-aware post-detection layer for OFDM-based integrated sensing and communication. The novelty is a private check-frame validation method that treats each public-frame detection as untrusted and verifies it using motion, waveform, and echo-power consistency. The method first applies standard range-Doppler processing and public-frame CFAR detection, then validates each candidate using temporal motion consistency across frames, a private sensing frame known only to the transmitter and receiver, and a private-to-public echo-power ratio test. Monte Carlo simulations include real targets, clutter, multipath, noise, residual carrier frequency offset, and four spoofing cases. Results show that the proposed method preserves real target detection while improving fake-target rejection from 0% to 82.5% for public-only smooth spoofing and from 0% to 94.5% for jumping spoofing. These results show that private waveform verification can add sensing-integrity validation to OFDM-based 6G sensing systems.

Temporal Windowing for Cross-Session Wi-Fi CSI Presence Detection

Andrei Țaga, Răzvan Crăciunescu, Răzvan Mihai, Mihnea Udrea, Constantin Pintilie

2026 16th International Conference on Communications (COMM) · 2026

Conference

Opportunistic Radar for the Automotive Industry

Abstract

Wi-Fi Channel State Information (CSI) enables device-free presence sensing, but temporal windowing can strongly affect cross-session detection performance. In this work, we analyze binary presence detection under a strict leave-one-session-out (LOSO) protocol using four continuous-stream CSI acquisitions collected under baseline, small-shift, non-line-of-sight (NLoS), and cross-room conditions. From the same recorded streams, we generate offline re-windowed datasets with temporal window lengths W ∈ {4, 8, 16} and hop sizes H ∈ {1, 2, 4}. Each sample is represented by CSI amplitude and its temporal difference, and all configurations are evaluated using the same lightweight convolutional neural network (CNN). We report accuracy, F1-score, and Matthews correlation coefficient (MCC). Results show that increasing the window length consistently improves cross-session robustness, while hop size has a comparatively smaller effect. The best overall performance is obtained for W = 16, indicating that window length is the dominant temporal design factor in this CSI-based presence detection pipeline.