Genetic Classification of Accented Speech from Audio Recordings of Spoken Nonsense Words

This thesis investigates the extent to which accented speech can be classified, based on their underlying genetic language class, from spoken nonsense words. Different models were compared; k-Nearest Neighbors (KNNs), Convolutional Neural Networks (CNNs), and Long-Short Term Memory Recursive Neural Networks (LSTM-RNNs). In addition, two different audio feature representation formats (Mel-Frequency Cepstral Coefficients (MFCCs) and the first two vowel formants), two taxonomic levels (family and genus) are explored. For the vowel formants representation (F1/F2), two ways of combining models are explored; early fusion combines all training data into one model and late fusion, which aggregates the prediction results per word. While it is inconclusive if language families can be used as a classification basis, it was found that the CNN using MFCCs was most effective in genera-based accent classification, achieving an accuracy of 60.49%, outperforming the baseline dummy classifier. These findings can thus serve as a baseline for language accent group classification in future works