International Journal of Intelligent Systems最新文献

{"title":"A Web-Based Intelligent Decision Support System Integrating Statistical Pre-Validation and AI-Assisted Interpretation for Multicriteria Decision Analysis","authors":"Anderson Gonçalves Portella,&nbsp;Miriam Karla Rocha,&nbsp;Robson Ferreira de Souza,&nbsp;Marcos dos Santos,&nbsp;Carlos Francisco Simões Gomes","doi":"10.1155/int/2388783","DOIUrl":"https://doi.org/10.1155/int/2388783","url":null,"abstract":"<p>Multicriteria decision analysis (MCDA) methods are widely used to evaluate alternatives under conflicting objectives, yet the statistical validity of input decision matrices is rarely verified prior to method application. Existing tools focus on executing MCDA algorithms without offering integrated statistical pre-validation or automated interpretation of results. This paper presents the design, implementation and evaluation of an open-source, web-based intelligent decision support system that integrates statistical pre-validation and AI-assisted interpretation for MCDA decision matrices. Developed following the Design Science Research methodology and implemented in Python with a Streamlit interface, the system encapsulates domain knowledge through six analytical modules: descriptive statistics, normality testing, correlation analysis, ANOVA with post hoc testing, linear regression with diagnostic assessment and Monte Carlo simulation. An AI-powered interpretation engine provides contextualised explanations and interactive consultation, encoding statistical reasoning that would otherwise require specialist expertise. The software is registered with the Brazilian National Institute of Industrial Property (INPI, BR512026000340-7) and publicly available on GitHub. The system was evaluated through a case study on intermodal transport selection for ethanol logistics in Brazil, involving four alternatives assessed across five criteria. The expert system identified strong intercriterion correlation (<i>r</i> = 0.99), validated ranking robustness via Monte Carlo simulation and generated AI-driven interpretations consistent with domain expert judgement. The proposed system advances the field of intelligent decision support by codifying statistical pre-validation expertise into an accessible, AI-augmented tool that bridges the gap between MCDA method execution and the rigorous statistical reasoning required for scientifically sound decision-making.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-09-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/2388783","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149048059","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"ENFCL: Self-Supervised Contrastive Learning Framework of Electric Network Frequency Representations for Generalizable Audio Tampering Detection","authors":"Chunyan Zeng,&nbsp;Fan Wang,&nbsp;Zhifeng Wang","doi":"10.1155/int/6478632","DOIUrl":"https://doi.org/10.1155/int/6478632","url":null,"abstract":"<p>Digital audio tampering detection is essential for preserving information authenticity, sustaining societal trust, and safeguarding public safety. However, most electric network frequency (ENF)–based detectors are trained in a fully supervised manner and therefore depend on large-scale labeled datasets, which are costly to obtain and often unavailable in real-world deployments. This heavy reliance on annotations limits representation discriminability and weakens generalization across unseen recording conditions and tampering patterns. To address these challenges, we propose ENF-based contrastive learning (ENFCL), a self-supervised framework that introduces contrastive pretraining for ENF-based audio tampering detection. ENFCL learns discriminative ENF representations from unlabeled audio by constructing diverse positive and negative sample pairs, thereby improving robustness and generalization without manual annotations. For pretraining, we design a multiscale encoder, multiscale ENF feature encoder with transformer (MSENFformer) to capture tampering cues at different resolutions. MSENFformer integrates three complementary subnetworks, zero-order phase deep residual network (ZPDR-Net), first-order phase deep representation network (FPDR-Net), and multiple-harmonics deep aggregation network (MHDA-Net) to model ENF characteristics across phase orders and harmonic components. A transformer module further aggregates global contextual information to enrich hierarchical feature representations. To mitigate the adaptation gap during fine-tuning, we introduce grouped convolution and dual-attention–based selective kernel network (GCDA-SKNet), which adaptively fuses multiscale features via dynamic selective kernels, enhancing discrimination for downstream classification. Experiments demonstrate that ENFCL achieves 94.00% accuracy on ENF-EDIT1, substantially outperforming existing state-of-the-art methods. Across ENF-EDIT1, ENF-EDIT2, and ENF-HBUT, ENFCL remains consistently strong under diverse data augmentation settings, confirming its robustness and generalization in practical audio tampering detection scenarios.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-09-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/6478632","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149048024","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"SpinalDenseXtNet: An Explainable Hybrid Deep Learning Framework for Mammographic Breast Cancer Classification","authors":"K. Sudha,&nbsp;G. Bhuvaneswari","doi":"10.1155/int/1034463","DOIUrl":"https://doi.org/10.1155/int/1034463","url":null,"abstract":"<p>Breast cancer is one of the primary causes of cancer-related deaths among women. Using mammography images for initial and precise findings is therefore vital. However, mammogram analysis is hindered by low contrast, noise, and high class similarity. This research presents an explainable and noise-resistant breast cancer recognition system utilizing a hybrid deep learning (DL) model called SpinalDenseXtNet. First, extended tuned adaptive frost filtering (Ext-AFF) is used to enhance mammographic images from the Mammographic Image Analysis Society (MIAS) and the Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM) datasets. PSPNet is employed for lesion segmentation. The segmented ROIs are then classified using the proposed SpinalDenseXtNet. Finally, Grad-CAM is used to visualize diagnostically relevant areas and enhance model interpretability. The proposed method’s performance is assessed using standard clinical metrics, thereby corroborating the efficiency and reliability of the proposed framework. The suggested model achieved 99.80% precision, 99.63% recall, 99.65% F1-score, 99.82% mean average precision (mAP), 98.48% Dice coefficient, and 99.71% accuracy.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-09-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/1034463","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148994235","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"DSM: Document Sentiment Map","authors":"Fatih Saglam,&nbsp;Burkay Genc,&nbsp;Hayri Sever","doi":"10.1155/int/4214284","DOIUrl":"https://doi.org/10.1155/int/4214284","url":null,"abstract":"<p>Sentiment analysis (SA) has traditionally been treated as a static text classification task, where a single polarity label is assigned to an entire document. However, this approach fails to capture the dynamic emotional transitions and sentiment fluctuations inherent in long-form texts. This study introduces a novel process-based framework, the Document Sentiment Map (DSM), which transforms SA into a longitudinal monitoring task. The DSM model contributes two major innovations: (i) the generation of Sentiment Oscillation Charts (SOC) to visualize emotional flow and (ii) the application of Shewhart Control Chart (SCC) rules to statistically interpret sentiment patterns. To establish linguistic reliability, we determined language-specific control limits using a massive Turkish language corpus of 2.12 million sentences. We demonstrate the model’s effectiveness through case studies on diverse datasets, including news articles and children’s literature. The results show that DSM successfully identifies fine-grained emotional shifts such as dominant, distinguishable, and powerful fluctuations that are invisible to traditional classifiers. By providing an interpretable and diagnostic view of sentiment, DSM offers a systematic solution for content auditing and explainable Natural Language Processing (NLP) applications. The DSM model shifts SA from being confined to a text classification problem to a process-based perspective, focusing on the process rather than just the outcome.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-09-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/4214284","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148994236","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A Fuzzy Choquet Integral–Based Hybrid Deep Learning Framework for Robust Multiclass Alzheimer’s Detection","authors":"Md. Mahmudul Hasan Sabbir,&nbsp;Nusrat Jahan Ritu,&nbsp;Md. Alamin Talukder,&nbsp;Mohsin Kazi,&nbsp;Mehnaz Tabassum","doi":"10.1155/int/4438827","DOIUrl":"https://doi.org/10.1155/int/4438827","url":null,"abstract":"<p>Reliable multiclass staging of Alzheimer’s disease from brain MRI remains challenging because adjacent disease stages often present subtle, heterogeneous, and visually overlapping structural patterns. Single deep-learning backbones may therefore produce unstable predictions, especially under class imbalance and borderline stage separation. To address this limitation, this study proposes FCI-ResNet, an interaction-aware fuzzy decision-fusion framework for four-class Alzheimer’s MRI classification. The framework integrates three ImageNet-pretrained ResNetV2 backbones, ResNet50V2, ResNet101V2, and ResNet152V2, as complementary feature learners and fuses their class-probability outputs using the fuzzy Choquet integral (FCI). Unlike voting, averaging, or distance-based fusion, the proposed FCI strategy models both individual learner reliability and interlearner dependency, enabling more principled decision fusion for uncertain multiclass predictions. A controlled experimental protocol was adopted on the Mendeley Alzheimer MRI dataset, including dataset verification, split-before-preprocessing, ROI-focused preprocessing, train-only augmentation, validation-only FCI optimization, isolated test-set evaluation, and class-imbalance–aware assessment. On the Mendeley test set, FCI-ResNet achieved 97.81% accuracy, 97.82% precision, 97.81% recall, 97.81% F1-score, 96.93% macro F1-score, 96.40% MCC, and 99.86% AUC, outperforming individual ResNetV2 backbones and the FHD comparator. External validation on the independent OASIS dataset further demonstrated strong generalization, achieving 99.31% accuracy, 99.47% precision, 99.47% recall, 99.31% F1-score, 99.47% macro F1-score, 99.00% MCC, and 99.98% AUC. Bootstrap confidence intervals, McNemar significance testing, discordant-pairs analysis, ablation study, confusion matrix evaluation, computational complexity analysis, and Grad-CAM visualization collectively support the robustness, reliability, interpretability, and practical relevance of the proposed framework. Overall, FCI-ResNet provides an externally validated and statistically supported MRI-based decision-fusion framework for robust multiclass Alzheimer’s disease staging.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-09-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/4438827","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148994005","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Digital Twin Fault Diagnosis Based on Convex Optimization for Industrial Gearbox Equipment","authors":"Suo Zhang,&nbsp;Xuelin Du,&nbsp;Xiaohu Wang,&nbsp;Caibiao Chen,&nbsp;Shuai Cao","doi":"10.1155/int/1016537","DOIUrl":"https://doi.org/10.1155/int/1016537","url":null,"abstract":"<p>This paper proposes a convex optimization digital twin (CODT) framework for fault diagnosis of intelligent manufacturing equipment, aiming to address the limitations of insufficient labeled samples, poor model interpretability, and low environmental adaptability in traditional data-driven diagnostic methods. The CODT method integrates convex optimization into a digital twin architecture to realize collaborative data fusion, physical constraint modeling, and adaptive decision-making. Multisource virtual–physical data are fused to construct high-dimensional fault features, and the fault diagnosis task is formulated as a constrained convex optimization problem. By combining cross-entropy loss with L₂ regularization and adopting a hybrid interior-point–gradient descent algorithm, globally optimal diagnostic solutions are obtained through dynamic virtual model updating. Experimental results on industrial gearbox datasets validate the performance of the proposed CODT framework. The method achieves a diagnostic accuracy of 98.5% and an F1 score of 98.2%, enabling early fault warning up to 50 time steps ahead. Comparative experiments with SVM, random forest, KNN, and GMM demonstrate that CODT provides more stable diagnostic confidence and lower detection latency. Robustness tests under noisy interference, cross-device scenarios, and complex operating conditions confirm the framework’s favorable generalization ability. Specifically, the accuracy degradation is only 2.5% under high-intensity noise, and the cross-scene accuracy fluctuation remains within 1%. Despite higher training overhead caused by convex optimization iterations and virtual–physical mapping, CODT improves diagnostic stability, interpretability, and proactive warning capability, offering a practical solution for industrial equipment health monitoring.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-09-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/1016537","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148943079","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Statistical and Spatiotemporal Hand Gesture Features for Sign Language Recognition Using the Leap Motion Sensor","authors":"Jordan J. Bird","doi":"10.1155/int/1930920","DOIUrl":"https://doi.org/10.1155/int/1930920","url":null,"abstract":"<p>Improvements in automatic sign language recognition (SLR) will lead to more enabling smart environments through digital technology to support good health and well-being. Many state-of-the-art approaches to SLR focus on the classification of static hand gestures or resource-intensive image processing. The contributions of this study are to explore optimal combinations of statistical features for an effective low-cost approach. Some signs are temporal, reflecting in many of the dynamic gestures present, which is not often considered within the state of the art. This work considers the problem of signed gesture recognition in terms of how dynamic gestures change during delivery and how different types of features affect the classification ability of a machine learning model. Eighteen common British Sign Language gestures recorded via a Leap Motion Controller sensor provide a complex classification problem, and two sets of statistical spatiotemporal attributes are extracted from 0.6-s time windows. The ANOVA <i>F</i>-scores and the <i>p</i> values show that many of the features are statistically significant. A total of 146 individual machine learning models are trained after dimensionality reduction, and the results show that the best feature selection approach is to combine statistical and spatiotemporal features, with the most significant 240 features of each type leading to a mean 86.75% K-fold accuracy. The results show that classification metrics always improve when combining the two types of features. Within a wider comparison, it was found that histogram gradient boosting could outperform the random forest with a mean accuracy of 88.089%; however, it had a much higher inference time.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-09-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/1930920","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148942582","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Real-Time Segmentation of B-Mode Ultrasound Images for Muscle Strain Using Deep Learning","authors":"Beiqi Li,&nbsp;Hongjie Zhao,&nbsp;Le Zhu,&nbsp;Li Zhu,&nbsp;Zihan Tang,&nbsp;Jing Zhou,&nbsp;Zhidong Xue","doi":"10.1155/int/6264568","DOIUrl":"https://doi.org/10.1155/int/6264568","url":null,"abstract":"<div>\u0000 \u0000 <section>\u0000 \u0000 <h3> Objective</h3>\u0000 \u0000 <p>The diagnosis of muscle strain using B-mode ultrasound is highly dependent on the sonographer’s experience, which can lead to diagnostic errors and prolonged examination times. This study aims to develop a real-time deep learning algorithm for the segmentation of muscle strain in B-mode ultrasound images, designed to assist sonographers by enhancing diagnostic accuracy and efficiency.</p>\u0000 </section>\u0000 \u0000 <section>\u0000 \u0000 <h3> Methods</h3>\u0000 \u0000 <p>A novel real-time instance segmentation model was proposed based on the YOLACT architecture. Key innovations include the integration of a VoVNet backbone for efficient feature extraction, an anchor-free prediction head incorporating multireceptive fields and a layer attention mechanism, and the application of deformable convolutions for mask coefficient prediction. The model was guided by a composite loss function, featuring an improved focal loss with soft labels to align classification and localization quality, and a distribution focal loss to enhance robustness against ambiguous annotations. The algorithm was trained and validated on a custom dataset of 1516 B-mode ultrasound images from 183 patients.</p>\u0000 </section>\u0000 \u0000 <section>\u0000 \u0000 <h3> Results</h3>\u0000 \u0000 <p>The proposed algorithm achieved state-of-the-art performance, yielding a bounding box AP50 of 81.8 and a mask AP50 of 81.4. Crucially, it demonstrated real-time capability with a processing speed of 40.9 frames per second (fps) on a GeForce RTX 4090 GPU. The model’s performance surpassed that of standard object detection and segmentation benchmarks. Ablation studies systematically validated the significant contribution of each methodological innovation to the final performance.</p>\u0000 </section>\u0000 \u0000 <section>\u0000 \u0000 <h3> Conclusion</h3>\u0000 \u0000 <p>This study successfully developed an efficient and accurate algorithm for the real-time segmentation of muscle strain in ultrasound images. Its high speed and accuracy demonstrate its potential as a valuable computer-aided diagnostic tool to support sonographers in clinical practice, ultimately aiming to improve the utilization of medical resources and enhance patient care.</p>\u0000 </section>\u0000 </div>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-09-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/6264568","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148942583","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"To Enhance the Prediction Accuracy of Carbon Trading Price: A Novel Model Combining Time–Frequency Decomposition and Hybrid Attention Mechanism","authors":"Haochu Cui,&nbsp;Yan Sun,&nbsp;Peiyue Cheng","doi":"10.1155/int/6531515","DOIUrl":"https://doi.org/10.1155/int/6531515","url":null,"abstract":"<p>Nowadays, carbon trading policy is regarded as an effective method to control carbon emissions. Accurate prediction of carbon trading price can benefit both governments and enterprises. To enhance the accuracy of prediction, this paper proposes a novel three-stage carbon trading price prediction model combining time–frequency decomposition and hybrid attention mechanism. In the decomposition stage, the historical price is decomposed into trend and seasonal components using multiwindow moving average and into high- and low-frequency components using fast Fourier transform. In the integration stage, a linear model predicts each component, and a hybrid attention mechanism is designed to calculate their dynamic weights to combine components to obtain distinct time-domain and frequency-domain prediction results. The fusion stage uses a linear network to integrate these intermediate results to accomplish the final carbon trading price prediction. Empirical data from Guangdong and Hubei are deployed to verify the proposed model. Comparative experiments demonstrate the proposed model reduces MAE by 1.1%–12.9% and MSE by 10.5%–13.1% for single-step forecasting compared to the best-performing baseline on each dataset, while maintaining competitive performance across most multistep predictions. Under accurate carbon trading price predictions, the governments can adjust quotas early to stabilize markets. Enterprises can also optimize allowance purchases or green investments based on price trends, enhancing emission reduction.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-09-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/6531515","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148942809","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Resource-Constrained Keyword Spotting With Quantized CNNs: Implementation Strategies and Performance Trade-Offs","authors":"Patrik Medur,&nbsp;Mart Lubbers,&nbsp;Goran Mauša","doi":"10.1155/int/2317323","DOIUrl":"https://doi.org/10.1155/int/2317323","url":null,"abstract":"<p>Resource-constrained embedded devices require specialized optimization strategies to achieve reliable real-time keyword spotting performance. This work presents a comprehensive implementation framework for deploying quantized convolutional neural networks on the ESP32-S3-EYE development board, achieving practical capabilities within severe memory and computational limitations. The implementation utilizes Google’s Speech Commands Dataset V0.02 with MFCC preprocessing, following the standard 12-class evaluation convention, 10 target keywords together with 2 auxiliary classes to ensure comparability with established KWS benchmarks. The deployed model achieves 89.58% accuracy in laboratory conditions with a quantized TensorFlow Lite (LiteRT) memory of 0.68 MB, enabling deployment on edge devices with limited memory resources. A dual-core system architecture implements FreeRTOS-based task distribution, dedicating one core to continuous audio capture, while the second handles feature extraction and inference operations. Voice activity detection with adaptive noise floor monitoring triggers processing when speech exceeds background noise levels. Comprehensive validation reveals substantial performance differences between controlled and real-world conditions. Testing with 13 non-native English speakers demonstrates 81.82% recognition accuracy. Synthetic speech evaluation using 10 AI voices achieves 90.06% accuracy, confirming that acoustic consistency significantly impacts recognition performance. The tiny machine learning framework enables practical keyword spotting on resource-constrained embedded systems for IoT applications.</p>","PeriodicalId":14089,"journal":{"name":"International Journal of Intelligent Systems","volume":"2026 1","pages":""},"PeriodicalIF":4.8,"publicationDate":"2026-08-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1155/int/2317323","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148943740","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}