Óscar Cánovas

This article introduces an automated system designed to classify teacher interventions in classrooms where Student Response Systems (SRS) are actively used. SRS tools, increasingly common in higher education, are valued for promoting more active learning and student engagement through real-time questioning and feedback cycles. Our methodology processes raw audio recordings, employing automatic speech recognition (ASR) and speaker diarization to generate transcripts and to derive several quantitative features. These transcripts serve as input for a series of advanced BERT-based models, which progressively leverage textual content, conversational context, and additional features to classify interventions into predefined COPUS-based pedagogical categories. A key contribution of our work is the empirical demonstration that hybrid models, fusing textual content with paralinguistic cues, can improve the performance of highly-optimized text-only models. This fusion provides a complementary source of information, and we couple our analysis with Explainable AI (xAI) techniques to clarify the distinct influence of these new features. Crucially, Explainable Artificial Intelligence (xAI) techniques, particularly SHAP, are applied to elucidate how these models fuse textual cues with paralinguistic elements (e.g., speaker ratios, silence patterns), revealing their differential contributions to the classification process. Ultimately, this study demonstrates a viable approach for creating accurate, interpretable AI systems that offer teachers feedback on their instructional patterns.

This systematic review synthesizes 82 peer-reviewed studies published between 2014 and 2024 on the use of audio features in educational research. We define audio features as descriptors extracted from audio recordings of educational interactions, including low-level acoustic signals (e.g., pitch and MFCCs), speaker-based metrics (e.g., talk-time and participant ratios), and linguistic indicators derived from transcriptions. Our analysis contributes to the field in three key ways: (1) it offers targeted mapping of how audio features are extracted, processed, and functionally applied within educational contexts, covering a wide range of use cases from behavior analysis to instructional feedback; (2) it diagnoses recurrent limitations that restrict pedagogical impact, including the scarcity of actionable feedback, low model interpretability, fragmented datasets, and limited attention to privacy; (3) it proposes actionable directions for future research, including the release of standardized, anonymized feature-level datasets, the co-design of feedback systems involving pedagogical experts, and the integration of fine-tuned generative AI to translate complex analytics into accessible, contextualized recommendations for teachers and learners. While current research demonstrates significant technical progress, its educational potential is yet to be translated into real-world educational impact. We argue that unlocking this potential requires shifting from isolated technical achievements to ethically grounded pedagogical implementations.

The increasing deployment of Internet of Things (IoT) devices in fog computing environments has amplified the need for efficient and scalable network threat detection. Traditional rule-based intrusion detection systems (IDSs), such as Suricata, struggle to keep pace with modern attack techniques and high-throughput network conditions, especially when facing unknown or zero-day threats. This paper presents a real-time anomaly detection architecture that leverages semi-supervised machine learning and GPU-accelerated inference using NVIDIA's Morpheus and Triton frameworks. Our approach operates at the fog level, enabling fast, localized detection without the need for labeled attack data. We conduct a performance comparison with Suricata, showing that the proposed system can sustain traffic analysis at nearly 19 Gbps—far surpassing the capabilities of traditional IDSs—while maintaining effective detection accuracy. Experimental evaluation using the CTU-IoT-23 dataset demonstrates the system's robustness and suitability for real-world fog-based deployments.

This article presents findings derived from the implementation of different modes of using an interactive response system in a Computer Engineering Degree course. The research focuses on the use of the Wooclap platform, which allows posing questions to students for real-time responses using their mobile devices. The designed experience addresses two modalities of platform use: one without competition and another with competition based on points. The methodology involves two distinct groups of students responding to the same questions, led by the same teacher in separate tests, using different platform modes. This quasi-experimental approach is based on data collection regarding student performance (correct responses to questions) and the overall level of classroom interaction during the tests. Furthermore, an audio recording analysis system supported by artificial intelligence is utilized to characterize the participation of the various actors involved.To further the analysis, transcriptions of the teacher’s oral interventions are examined, investigating the typology of these interventions and evaluating if the quantity and distribution of these types vary between the competition mode and the non-competition mode. The results reveal distinctive patterns in the level of student interaction during the tests, depending on the mode used, as well as in the type of interventions made by the teacher. Despite these variations, the results indicate that the mode of use does not have a significant influence on the performance achieved by the students.

Using automated models to analyze classroom discourse is a valuable tool for educators to improve their teaching methods. In this paper, we focus on exploring alternatives to ensure the generalizability of models for identifying teaching practices across diverse teaching contexts. Our proposal utilizes artificial intelligence to analyze audio recordings of classroom activities. By leveraging deep learning for speaker diarization and traditional machine learning algorithms for classifying teaching practices, we extract features from the audio diarization using a processing pipeline to provide detailed insights into teaching dynamics. These features enable the classification of three distinct teaching practices: lectures, group discussions, and the use of audience response systems. Our findings demonstrate that these features effectively capture the nuances of teacher-student interactions, allowing for a refined analysis of teaching styles. To enhance the robustness and generalizability of our model, we explore various pipelines for audio processing, evaluating its performance across different contexts with diverse teachers and students. By comparing these practices and their associated features, we illustrate how AI-driven tools can support teachers in reflecting on and improving their teaching strategies.

Fog computing, an extension of cloud computing, places computational resources closer to data sources, enhancing network efficiency and security. This paper addresses the challenges posed by cybersecurity threats within fog computing environments, specifically in managing vast volumes of network traffic from IoT devices which are vulnerable to cyberattacks. We propose a novel architecture for real-time anomaly detection that integrates fog and cloud computing layers. This two-tiered system utilizes machine learning and deep learning models for dynamic and adaptive threat detection. By leveraging the computational power of modern GPUs, the system analyzes network traffic at speeds up to 10 Gbps, enabling the detection of anomalies before they can cause significant damage. We introduce a distributed anomaly supervisor that operates at the fog level to provide real-time analysis and immediate response to potential threats, while cloud-level operations facilitate a comprehensive overview and coordination. The effectiveness of this architecture is demonstrated through extensive experimentation, proving its capability to manage high-speed data transfers and improve security measures across fog networks.

Teachers require reliable feedback to refine their skills and to analyze their methods in the classroom. However, it can be challenging for teachers to be aware of the characterization and distribution of classroom activities while simultaneously delivering lectures. Traditional methods of achieving expertise through deliberate practice under the guidance of a human coach are not suitable, due to the long turnover time required for training coaches, observing classrooms, and coding activities. To address these challenges, automated approaches using artificial intelligence (AI) techniques to analyze audio recordings have been proposed to infer the climate in classrooms, model teacher discourse, and classify teaching activities. While these approaches have shown promising results inferring activities, there is a need for additional proposals that support tools and analytics enabling teachers to reflect on their practice and track their progress. In this paper, we present a novel framework that leverages deep learning for speaker diarization and machine learning algorithms for the classification of teaching practices and the analysis of different teaching styles. Our approach utilizes several non-verbal discursive features to provide informative insights. Specifically, we have defined 12 different features that are employed to classify up to three distinct practices: lecture, group discussion, and the use of audience response systems. We show that those features are also informative to analyze the behavior of the teachers for each teaching practice.