PUB - Publikationen an der Universität Bielefeld

This is the first machine learning based analysis of oral vowels produced in semi-spontaneous Pakistani Punjabi. Our investigation is based on lower frequency characteristics (f0, F1, F2, F3) and vowel duration. We use random forest classification to investigate the prediction of Punjabi vowels based on these acoustic features. For allophonic variation, we use a confusion matrix to identify the false positive and false negative vowel predictions. Our data indicate that frequency characteristics do not reliably classify different vowels of the same class (e.g., front or high). This results in varying rates of successful predictions in the classification of vowels. The analysis of feature importance shows that F1 and F2 play the most important role in the classification of Punjabi vowels, followed by duration. The influence of F3 and f0 is negligible in this regard. The misclassification of vowels in the same group highlights the phonemic vs. allophonic variation in Pakistani Punjabi.