Investigating syllabic prominence with Conditional Random Fields and Latent-Dynamic Conditional Random Fields

The present study performs an investigation on several issues concerning the automatic detection of prominences. Its aim is to offer a better understanding of the prominence phenomenon in order to be able to improve existent prominence detection systems. The study is threefold: first, the presence of hidden dynamics in the sequence of prominent and non-prominent syllables is tested by comparing results obtained with CRFs and LDCRFs. Second, the size of the context to be taken into account when determining prominence was examined and third, a new set of features was investigated. The obtained results show that LDCRFs systematically outperform CRFs, that a context of three syllables is generally sufficient for prominence detection and that syllable length is a useful feature to include. Also, new features concerning pitch movements we introduced can substitute adequately heuristic measures used in previous works.