<![CDATA[This study presents the development of a speech training application for children with Down Syndrome, utilizing various machine learning algorithms to evaluate pronunciation accuracy. The study compares four widely used algorithms: K Nearest Neighbors (KNN), Naive Bayes, Support Vector Machine (SVM), and Random Forest, in classifying speech accuracy based on Mel Frequency Cepstral Coefficients (MFCC) features. Additionally, the impact of balancing the dataset using Synthetic Minority Oversampling Technique (SMOTE) is evaluated through 5 fold cross validation. The experimental results demonstrate that Random Forest achieves the highest performance with an overall accuracy of 85.35%, precision of 85.39%, recall of 41.79%, F1 score of 56.19%, and a cross validation accuracy of 85.01%, proving its capability to handle complex speech data. In comparison, KNN attains a cross validation accuracy of 80.54%, balancing precision and recall. Naive Bayes yields an overall accuracy of 80.09% (77.83% in cross validation) and a precision of 78.63%, constrained by its feature independence assumption. Meanwhile, SVM achieves a 77.58% cross validation accuracy but records a recall of 0.00% on the baseline dataset, indicating a failure to detect the minority class of poor pronunciation. The application of SMOTE effectively mitigates class imbalance and enhances recall across all models, particularly for KNN and Naive Bayes. Feature importance analysis confirms that lower order MFCC coefficients contribute significantly to classification accuracy. These quantitative findings underscore the viability of implementing Random Forest into speech therapy systems to deliver objective, robust, and accessible pronunciation assessments for children with Down Syndrome.]]>
Copyrights © 2025
