Smart assistants and smart microphones can contribute to detecting verbal abuse by analyzing speech-to-text contents. In this paper, we compare different large language models (LLMs) for detecting verbal abuse and propose a framework that first identifies emotion from short audio conversations by extracting Mel Frequency Cepstral Coefficient (MFCC) and Mel Spectrogram (MEL) features. It then utilizes transfer learning with a fully connected neural network incorporating an attention mechanism and SBERT encoding. To demonstrate the efficacy of the proposed framework, we prepared a custom dataset containing instances of verbal abuse. Evaluation results show that our framework is lightweight and achieves commendable accuracy compared to the existing LLM models.

With the increasing adoption of smart assistants, voice-enabled interactions have the potential to transform traditional classrooms into intelligent learning environments. A voice-enabled smart assistant can be designed to recognize context from student questions and brief conversations, enabling efficient query handling with minimal human intervention. This allows instructors to focus more on students who require personalized assistance, improving the overall learning experience. However, accurately identifying context remains a challenge due to the complexity and variability of student interactions. Therefore, in this paper, we present a comparative study of transformer models for context recognition (CoRe) in K-12 smart classrooms. We curated a custom dataset comprising short commands and conversations related to day-to-day classroom operations and trained multiple transformer models to identify the best-performing model for recognizing context. To ensure data quality and support data-driven decision-making, we performed topic modeling. Our evaluation results show that a transformer model enhanced with an attention mechanism outperforms existing transformer models while maintaining low computational costs, making it a viable solution for real-world smart classroom applications.

As smart IoT devices become increasingly common in our homes, there is a growing demand for seamless automation and synchronization between these devices. A key challenge in achieving this automation is accurately identifying and grouping related sensors, which are essential for generating automated operational policies to control the actuators. However, this process is often hindered by anomalous data in sensor readings, which can obscure the sensor relationships identified during the grouping process. These anomalies disrupt the accuracy of the sensor groupings and, as a result, compromise the effectiveness and reliability of the automation policies generated for managing smart environments. In this paper, we introduce SEAD, a novel approach for detecting anomalies by first calculating the total number of sensor events within a specified time window, followed by the use of an unsupervised learning method for anomaly detection. We evaluate the effectiveness of this method by leveraging existing sensor inference techniques and testing it on three custom datasets and one public dataset. Our experimental results show that applying SEAD to remove anomalies improves the quality of sensor groupings for smart home automation.

Modern smart homes comprise a large amount of sensors and actuators. Identifying sensor relationships can contribute to automating operational policies of the actuators. However, in a multi-resident smart home, it is difficult to identify sensor relationships due to a variety of simultaneous sensor events. In this paper, we propose a novel two-step approach, which initially extracts features from time series data generated by sensor events to cluster related sensors, and then identifies how each sensor is related to other sensors revealing their physical proximity and enabling a sensor's ability to participate in multiple sensor groups. Experimental results show that our approach performs well in multi-resident homes even if the sensor events are not equally distributed.

Information sharing among agents to jointly solve problems is challenging for multi-agent reinforcement learning algorithms (MARL) in smart environments. In this paper, we present a novel information sharing approach for MARL, which introduces a Team Information Matrix (TIM) that integrates scenario-independent spatial and environmental information combined with the agent's local observations, augmenting both individual agent's performance and global awareness during the MARL learning. To evaluate this approach, we conducted experiments on three multi-agent scenarios of varying difficulty levels implemented in Unity ML-Agents Toolkit. Experimental results show that the agents utilizing our TIM-Shared variation outperformed those using decentralized MARL and achieved comparable performance to agents employing centralized MARL.