Image-Based Texture Analysis Of Vowel Spectrograms In Sylheti Using Random Forest Classifier

This study presents a novel approach to vowel classification in Sylheti using image-based textural analysis. Traditional methods, which often rely on formant frequencies or Mel-frequency cepstral coefficients (MFCCs), have been supplemented in this research with an automated image-processing technique that quantifies vowel characteristics through texture features. The study employs FOS (First Order Statistics), GLCM (Gray Level Co-occurrence Matrix), GLSZM (Gray Level Size Zone Matrix), GLRLM (Gray Level Run Length Matrix), and GLDM (Gray Level Dependence Matrix) to capture the unique acoustic properties of five Sylheti vowels [a, ε, i, o, and u]. The Random Forest classifier, applied to these features, demonstrates impressive accuracy, with an average AUC (Area Under the Curve) of 0.94, showing the robustness of the proposed method. The results reveal significant insights into the articulatory and acoustic distinctions among the vowels, highlighting the potential of textural analysis in advancing speech recognition technologies. Future work will aim to refine the feature extraction process and explore more sophisticated machine learning algorithms to further improve classification accuracy.