An approach to lexical stress detection from transcribed continuous speech using acoustic features

This paper presents a first approach to the unsupervised learning and prediction of primary lexical stress starting from continuous speech data and its orthographic transcript. The approach is intended to be used in the development of text-to-speech synthesis systems for under-resourced languages. Our method is based on syllable nuclei approximation and stress detection using simple acoustic features. The evaluation is performed on 3.5 hours of speech uttered by a Romanian female speaker and results show an accuracy of 47.20% at word level and 58.61% at syllable level.