Assessing reading fluency in elementary grades: A machine learning approach

Texto completo

This study compares eleven widely used machine learning algorithms to identify the most accurate and comprehensive method for assessing children's reading fluency. Our fluency framework integrates three key dimensions: accuracy in word decoding, reading speed (words per minute), and prosody, which captures appropriate pausing and intonation during reading. Audio recordings from 2nd and 3rd grade students across 144 Brazilian schools were transcribed using the advanced Whisper ASR system, enabling automated extraction of fluency features. The research objective was to determine which algorithm best predicts fluency, considering diverse evaluation setups including binary classification (fluent versus non-fluent), multiclass classification (differentiated fluency levels), and regression analysis to estimate continuous fluency scores. Among the eleven models evaluated, Logistic Regression achieved the highest overall performance in the classification experiments, demonstrating superior precision and accuracy. Ensemble methods such as Gradient Boosting and Random Forest also yielded robust results, particularly in regression analyses where they effectively captured the variability in expert fluency ratings. Notably, reading speed emerged as the most critical indicator of fluency across all experiments, consistently outweighing contributions from accuracy and prosody. Nevertheless, prosody maintained significant importance, especially when fluency was modeled as a continuous variable, emphasizing the value of expressive reading. These findings suggest that while high accuracy in word decoding is fundamental, interventions that enhance reading speed and incorporate prosodic features can offer more nuanced support for reading development. This work provides educators with insights into scalable, automated, and cost-effective approaches to monitoring and improving reading fluency.