Automated phonetic transcription of Croatian folklore genres using supervised machine learning

Abstract

This paper aims to detect the possibilities of automatic text transcription for the purpose of preparing a corpus for further natural language processing analysis. The corpus contains various Croatian folklore genres. The transcription goal is to have one character represent one phoneme and remove spaces between accentuated and non-accentuated words. This knowledge independent system is trained using supervised learning methods and applied to the rest of the corpus using classifiers such as the naïve Bayes, k-nearest neighbour, support vector machine and others. The results are compared to a human-annotated sample to determine accuracy.