The common ToBI system of transcription assumes a sequential model of prosody. Many linguists argue for a tree structure explaining the synchronization and interaction among prosodic units. Could tree grammars, used previously in syntax-based language modeling, be used to model prosodic trees? We present a method of converting sequential transcripts into trees, and then demonstrate that modeling trees rather than sequences of prosodic tags results in lower perplexity as well as lower error rates when classifying pitch accents and boundaries on the Boston University Radio News Corpus. This finding could benefit areas like speech synthesis, speech understanding, and pronunciation evaluation.
Bibliographic reference. Tepperman, Joseph / Narayanan, Shrikanth S. (2008): "Tree grammars as models of prosodic structure", In INTERSPEECH-2008, 2286-2289.