Selected work

Representative Papers

Soroll-IA: A Weakly Labeled Audio Dataset for Real-World Industrial Port Monitoring

Jun 24, 2026

This work addresses the lack of weakly labeled audio datasets tailored to real-world industrial port environments, which has hindered the development of robust sound event detection and acoustic analysis. To bridge this gap, we introduce and publicly release the first weakly annotated audio dataset for industrial ports, comprising approximately 22 hours of audio (7,396 segments) recorded at fixed sensor nodes in the Port of Valencia, Spain, capturing 26 representative sound source classes. Two versions of weak labels, derived from expert consensus, are provided. We establish benchmark tasks that account for challenges such as high ambient noise, long-distance recording, and overlapping events, and evaluate performance using CNN14 for high-accuracy audio tagging and MobileNetV2 for edge-compatible real-time classification. This dataset serves as a valuable benchmark for weakly supervised sound event detection, audio tagging, and machine learning under low-resource conditions.

0 citationsRead paper

ORICF -- Open Robotics Inference and Control Framework

May 10, 2026

This work addresses the challenges of deploying advanced AI models on robots, which are often constrained by high computational overhead, latency, and energy consumption. The authors propose a modular, declarative, and model-agnostic multimodal inference framework that enables flexible composition of models, hardware, and data pipelines through lightweight YAML configurations, supporting edge offloading without code modifications. Built upon ROS 2, the system integrates components such as automatic speech recognition (ASR), large language models (LLMs), and convolutional neural networks (CNNs), leveraging I/O adapters, pluggable inference backends, and post-processing logic to ensure modularity and reproducibility. Experimental results demonstrate that, compared to fully onboard execution, the edge-deployed configuration reduces on-robot computational utilization by 83.16% and decreases energy consumption by 65.8%.

0 citationsRead paper

Metrics, KPIs, and Taxonomy for Data Valuation and Monetisation - Internal Processes Perspective

Dec 11, 2025

In data-driven economies, organizations lack systematic frameworks for evaluating and managing data value within internal business processes. To address this gap, this study develops a comprehensive data value assessment framework grounded in the Balanced Scorecard’s internal process perspective, integrating three interrelated dimensions: data quality, governance compliance, and operational efficiency. It introduces a novel, multi-layered taxonomy of data value—spanning technological, organizational, and regulatory dependencies—that resolves metric redundancy and establishes cross-dimensional conceptual linkages. Through systematic literature review, theoretical modeling, indicator clustering, and taxonomy design, the research produces a scalable, reusable data value metrics system. This system underpins standardized data valuation models and decision-support systems, offering both a methodological foundation and actionable implementation pathways for cross-sectoral data assetization. (149 words)

0 citationsRead paper

Threat Modeling for Enhancing Security of IoT Audio Classification Devices under a Secure Protocols Framework

Sep 18, 2025

To address security threats against resource-constrained IoT audio nodes processing highly sensitive voice data on-device, this paper proposes a cross-trust-domain defense-in-depth architecture. The approach integrates STRIDE threat modeling with attack tree analysis to design a secure boot chain supporting remote attestation and TLS 1.3 mutual authentication. It incorporates TPM-based PCR measurements, LUKS full-disk encryption, Kyber/Dilithium post-quantum hybrid encryption, signature-protected AI models, and tamper-evident sensors. Key innovations include one-time key release, firmware rollback protection, and end-to-end, lifecycle-wide data protection. Experimental evaluation demonstrates that the architecture effectively blocks known attack vectors, prevents activation of compromised or physically altered devices, and satisfies rigorous physical and logical security requirements. Security validation confirms robustness across both hardware and software attack surfaces.

0 citationsRead paper

Comprehensive Evaluation of CNN-Based Audio Tagging Models on Resource-Constrained Devices

Sep 17, 2025

Deploying CNN-based audio tagging models on resource-constrained edge devices (e.g., Raspberry Pi) faces challenges including high inference latency, thermal runaway, and poor long-term operational stability. This work systematically evaluates six model families—PANNs (1D/2D), ConvNeXt, MobileNetV3, and two newly designed lightweight architectures (CNN9 and CNN13)—under 24-hour continuous inference, measuring both accuracy and thermal behavior. We introduce, for the first time, a standardized long-term stability evaluation protocol and unify deployment via ONNX to ensure cross-platform compatibility and efficiency. Results show that CNN9 achieves the optimal trade-off among accuracy, latency, and power consumption: its average latency variation remains below 1.2% over 24 hours, with temperature rise stabilized at ≤15°C. ONNX conversion improves deployment portability and delivers up to 1.8× inference speedup. This study provides a reproducible, sustainable, and lightweight deployment framework for edge audio perception.

0 citationsRead paper

Recent publications

Latest Papers

Soroll-IA: A Weakly Labeled Audio Dataset for Real-World Industrial Port Monitoring

Jun 24, 2026

This work addresses the lack of weakly labeled audio datasets tailored to real-world industrial port environments, which has hindered the development of robust sound event detection and acoustic analysis. To bridge this gap, we introduce and publicly release the first weakly annotated audio dataset for industrial ports, comprising approximately 22 hours of audio (7,396 segments) recorded at fixed sensor nodes in the Port of Valencia, Spain, capturing 26 representative sound source classes. Two versions of weak labels, derived from expert consensus, are provided. We establish benchmark tasks that account for challenges such as high ambient noise, long-distance recording, and overlapping events, and evaluate performance using CNN14 for high-accuracy audio tagging and MobileNetV2 for edge-compatible real-time classification. This dataset serves as a valuable benchmark for weakly supervised sound event detection, audio tagging, and machine learning under low-resource conditions.

0 citationsRead paper

ORICF -- Open Robotics Inference and Control Framework

May 10, 2026

This work addresses the challenges of deploying advanced AI models on robots, which are often constrained by high computational overhead, latency, and energy consumption. The authors propose a modular, declarative, and model-agnostic multimodal inference framework that enables flexible composition of models, hardware, and data pipelines through lightweight YAML configurations, supporting edge offloading without code modifications. Built upon ROS 2, the system integrates components such as automatic speech recognition (ASR), large language models (LLMs), and convolutional neural networks (CNNs), leveraging I/O adapters, pluggable inference backends, and post-processing logic to ensure modularity and reproducibility. Experimental results demonstrate that, compared to fully onboard execution, the edge-deployed configuration reduces on-robot computational utilization by 83.16% and decreases energy consumption by 65.8%.

0 citationsRead paper

Metrics, KPIs, and Taxonomy for Data Valuation and Monetisation - Internal Processes Perspective

Dec 11, 2025

In data-driven economies, organizations lack systematic frameworks for evaluating and managing data value within internal business processes. To address this gap, this study develops a comprehensive data value assessment framework grounded in the Balanced Scorecard’s internal process perspective, integrating three interrelated dimensions: data quality, governance compliance, and operational efficiency. It introduces a novel, multi-layered taxonomy of data value—spanning technological, organizational, and regulatory dependencies—that resolves metric redundancy and establishes cross-dimensional conceptual linkages. Through systematic literature review, theoretical modeling, indicator clustering, and taxonomy design, the research produces a scalable, reusable data value metrics system. This system underpins standardized data valuation models and decision-support systems, offering both a methodological foundation and actionable implementation pathways for cross-sectoral data assetization. (149 words)

0 citationsRead paper

Threat Modeling for Enhancing Security of IoT Audio Classification Devices under a Secure Protocols Framework

Sep 18, 2025

To address security threats against resource-constrained IoT audio nodes processing highly sensitive voice data on-device, this paper proposes a cross-trust-domain defense-in-depth architecture. The approach integrates STRIDE threat modeling with attack tree analysis to design a secure boot chain supporting remote attestation and TLS 1.3 mutual authentication. It incorporates TPM-based PCR measurements, LUKS full-disk encryption, Kyber/Dilithium post-quantum hybrid encryption, signature-protected AI models, and tamper-evident sensors. Key innovations include one-time key release, firmware rollback protection, and end-to-end, lifecycle-wide data protection. Experimental evaluation demonstrates that the architecture effectively blocks known attack vectors, prevents activation of compromised or physically altered devices, and satisfies rigorous physical and logical security requirements. Security validation confirms robustness across both hardware and software attack surfaces.

0 citationsRead paper

Comprehensive Evaluation of CNN-Based Audio Tagging Models on Resource-Constrained Devices

Sep 17, 2025

Deploying CNN-based audio tagging models on resource-constrained edge devices (e.g., Raspberry Pi) faces challenges including high inference latency, thermal runaway, and poor long-term operational stability. This work systematically evaluates six model families—PANNs (1D/2D), ConvNeXt, MobileNetV3, and two newly designed lightweight architectures (CNN9 and CNN13)—under 24-hour continuous inference, measuring both accuracy and thermal behavior. We introduce, for the first time, a standardized long-term stability evaluation protocol and unify deployment via ONNX to ensure cross-platform compatibility and efficiency. Results show that CNN9 achieves the optimal trade-off among accuracy, latency, and power consumption: its average latency variation remains below 1.2% over 24 hours, with temperature rise stabilized at ≤15°C. ONNX conversion improves deployment portability and delivers up to 1.8× inference speedup. This study provides a reproducible, sustainable, and lightweight deployment framework for edge audio perception.

0 citationsRead paper