Mispronunciation Detection Using Feature Learning

This research describes a study into the use of feature learning approaches for mispronunciation detection. Mispronunciation detection is essential in speech recognition and language learning applications. Mispronunciation detection has traditionally depended on handcrafted features and rule-based algorithms, which frequently have poor generalisation abilities and demand a lot of manual work but recent advances in deep learning and feature learning have demonstrated promising results in enhancing the accuracy and robustness of mispronunciation detection systems. The study proposes an innovative method using Mel-frequency cepstral coefficients (MFCC) and SVM. Comparing TF-IDF and MFCC feature extraction, labelled audio recordings are categorised as accurate or incorrect pronunciation. Audio data is preprocessed, and MFCC features are extracted using the librosa library. SVM learns patterns between features and labels during training. Performance is evaluated on a common voice dataset, achieving 71% accuracy with MFCC and 70% with TF-IDF. Additional preprocessing and hyperparameter tuning result in 71% accuracy for TF-IDF. Overall, this research shows that MFCC feature extraction and an SVM classifier are effective tools for mispronunciation identification. By utilising the strength of feature learning on the Common Voice dataset, this research advances mispronunciation detection systems, ultimately enhancing language learning processes and facilitating the creation of more precise speech processing applications.