Speech Fluency Features for Robust Automatic Classification of L2 English Speech using a Small Dataset

This work presents an automatic fluency level classifier of L2 English speech. Using models such as Support Vector Machines, and a Multilayer Perceptron, comparing features of three previous works [3], [20], [12] such as speech rate, filled-pauses, phonation ratio, effective speech rate, and the well known Mel Frequency Cepstral Coefficients (MFCC). Furthermore, we utilize the Avalinguo dataset, which [12] proposed. It comprises 1424 audio samples of individuals speaking, categorized into three fluency levels: Basic, Intermediate, and Advanced. Additionally, we introduce two experiments that demonstrate the robustness of some of these features and models. Our findings indicate that SVM with Yu & Van Heuven [20] features provides the highest accuracy and demonstrates the most robustness among all tested models and features.

Paper: Google Drive