AI Model Scores College Spoken English With 95 Percent Accuracy

Spoken English has become a gatekeeping skill in higher education and the workplace, yet most college courses still rely on lecture-based instruction with few chances for structured speaking practice and almost no systematic evaluation. Instructors who do grade speaking performance face well-documented problems: the assessments are subjective, slow, and inconsistent, while students often feel anxious or shy about speaking in class, which erodes both confidence and learning outcomes. A new open-access study published in Discover Artificial Intelligence on 27 October 2026 proposes a technical answer to that gap: an Artificial Intelligence and Internet of Things based training and evaluation system, abbreviated IoT-TES, built around a hybrid machine learning model that the authors report achieving 95.47 percent accuracy in classifying spoken English performance.

The study, authored by Manli Jia, targets a recurring weakness in existing digital spoken English assessment tools. According to the paper, current systems lack adaptive evaluation mechanisms and struggle to handle diverse acoustic conditions and contextual learning environments, which undermines assessment reliability. Some systems focus narrowly on pronunciation rather than evaluating speaking skills comprehensively; others generalize poorly across different speakers and recording conditions, rely on inadequate features, produce uneven predictions, and offer little actionable feedback on how an individual learner might improve. The author argues that poor integration of advanced optimization techniques with ensemble learning has limited both predictive power and computational performance, motivating a more adaptable framework.

The technical core of the system is a model the paper calls AOA-RXGBoost, a combination of the Archimedes Optimization Algorithm with a hybrid of Random Forest and Extreme Gradient Boosting. The Archimedes Optimization Algorithm performs two jobs simultaneously: it selects the most discriminative subset of speech features and tunes the hyperparameters of the classification model, reducing feature redundancy and speeding convergence. Random Forest then supplies robust predictions through ensemble learning, aggregating many decision trees to reduce variance and resist overfitting, while XGBoost refines the results by training successive trees to minimize the residual errors of earlier ones. The paper describes the Random Forest stage as producing both class predictions and confidence scores, using out-of-bag samples and proximity measures between samples, and the XGBoost stage as applying regularization to control model complexity.

Before any classification happens, the system runs a speech processing pipeline designed for messy, real-world audio. Raw signals captured through IoT-enabled microphones are first cleaned with Wiener filtering, a classical noise reduction technique that models the observed signal as clean speech plus additive noise and estimates the clean spectrum by minimizing mean square error. Voice Activity Detection then separates speech from silence and noise by combining short-term energy and zero-crossing rate, and the signals are normalized for consistency. Feature extraction follows using mel-frequency cepstral coefficients, or MFCCs, a representation chosen because it mimics human auditory perception and captures spectral and temporal properties relevant to pronunciation, fluency, and prosody. The pipeline includes pre-emphasis, Hamming windowing, a fast Fourier transform, mapping onto the mel scale, a discrete cosine transform, and delta coefficients computed over a five-frame window to capture temporal dynamics.

The evaluation data come from a publicly available Kaggle dataset, the Spoken English Speech Evaluation Dataset, comprising 5,245 WAV audio recordings from 250 speakers collected across diverse environments including classrooms, laboratories, residential areas, and open spaces. Each audio example includes speaker-specific information and spoken characteristics such as MFCC, pitch, energy, pronunciation accuracy, fluency, prosody, comprehensiveness, and overall spoken English performance. The dataset was split into 4,196 training samples and 1,049 test samples in an 80:20 ratio. The paper frames the IoT architecture as three levels: microphones, mobile phones, and classroom sensors gather speech data across varied acoustic conditions, a network layer transmits it, and a processing unit runs the pre-processing and AI-based evaluation.

The reported results are the study’s headline claim. On the test set, the proposed model achieved 95.47 percent accuracy, 95.12 percent precision, 94.86 percent recall, and an F1-score of 94.99. The confusion matrix across four skill levels, Poor, Average, Good, and Excellent, shows diagonal values dominating each row, with most errors occurring between adjacent skill levels rather than distant ones, which the author interprets as the model reliably distinguishing clear-cut cases while showing minor confusion only at skill-level boundaries. The system was implemented in Python, and the code has been made publicly available through a GitHub repository to support reproducibility.

To contextualize those numbers, the author retrained several existing models on the same Kaggle dataset using identical preprocessing, feature extraction, and evaluation protocols, including a CNN with genetic algorithm, an LSTM model, DSM-RoBERTa, SNN-ConvLSTM, and an Improved DNN. The paper reports that AOA-RXGBoost achieved the highest values on all four metrics. The discussion section attributes the comparative weaknesses of the alternatives to specific causes: DSM-RoBERTa, despite strong contextual representation, fails to detect subtle differences in speech data; SNN-ConvLSTM performs acceptably but is affected by noise and inconsistencies; and the Improved DNN lacks optimal feature selection and can overfit on multidimensional MFCCs, leaving all of them less able to handle inter-speaker variability and different accents.

The study includes several robustness checks. A 5-fold cross-validation showed consistent performance with little variation among folds, with a small standard deviation and narrow confidence interval that the author reads as evidence of generalization. An ablation analysis showed incremental accuracy gains from Random Forest alone, to XGBoost alone, to the hybrid, with the full AOA-RXGBoost reaching the top value of 95.47. A sensitivity analysis of the AOA population size found the best classification performance at a population of 60, with performance declining slightly at both smaller and larger sizes. Exploratory analyses also documented why the preprocessing matters: environmental noise substantially affects MFCC feature quality and evaluation scores, recording devices differ markedly in prosodic quality, with studio microphones stable but smartphones, laptops, and a Bluetooth headset varying or degrading, and quiet rooms yield higher accuracy and fluency scores than noisy ones.

Beyond scoring, the system converts predictions into learner-specific feedback. Low fluency scores trigger suggestions to speak more slowly and maintain a consistent rhythm, while detected pronunciation errors generate articulation and word-stress recommendations. The paper describes this prediction-to-feedback mapping as enabling targeted pedagogical actions that support continuous improvement in spoken English proficiency, positioning the platform as a scalable solution for real-time feedback, personalized learning support, and adaptive evaluation across diverse educational environments.

The author is explicit about the study’s limitations, and they are substantial for anyone hoping to see this in classrooms tomorrow. MFCC features may not fully capture complex semantic and contextual speech characteristics, and the optimization plus ensemble learning process increases computational complexity, which could hinder deployment on resource-constrained edge devices. Critically, although the system is framed around IoT hardware, the model was developed and evaluated entirely on a pre-existing Kaggle dataset rather than on speech collected live through IoT-enabled microphones, mobile devices, and classroom sensors; validating the model on real-time IoT-collected classroom data is listed as future work, along with incorporating transformer-based speech embeddings and multimodal signals such as facial expressions and lip movements, extending the approach to multilingual assessment, and optimizing for low-latency edge deployment. Until those validations occur, the 95.47 percent figure describes performance on a curated public dataset, not a proven classroom deployment.

Subject of Research: An AI and IoT-based system for training and evaluating college spoken English using a hybrid machine learning model

Article Title: AI and IoT-based training and evaluation system for college spoken English

Article References: Jia, M. (2026). AI and IoT-based training and evaluation system for college spoken English. Discover Artificial Intelligence, 6(1), Article 1663. https://doi.org/10.1007/s44163-026-02294-w

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02294-w

Keywords: artificial intelligence, Internet of Things, spoken English assessment, machine learning, random forest, XGBoost, Archimedes Optimization Algorithm, MFCC, speech processing, Wiener filtering, personalized learning, educational technology