Improved Hybrid Tuning Mel Frequency Cepstral Coefficients with Ant Colony Optimization, and Long Short Term Memory on Speech Hoarseness Detection

Hoarseness speech detection through machine learning has been discussed quite extensively. However, not many people are trying to apply with different datasets and identify the type of algorithm that would be able to produce high accuracy, with the appropriate precision, recall, and F1-score. Two types of datasets are used in this study, including the Kaggle Speech dataset and the Saarbrucken Voice Dataset (SVD). The disadvantages of the Mel Frequency Cepstral Coefficient that affect the accuracy rate are overcome by using feature selection techniques, pitch features, and the selection of appropriate coefficients. From this technique, the accuracy rate has increased, especially using the selection of different coefficient parameters and the feature selection technique. Through this study, the increase in accuracy and increased performance metrics show the advantages of machine learning techniques in identifying hoarse and normal voices, especially in cancer patients.