Prosodic trees as a hierarchical representation of prosodic organization in French proved to be efficient for automatic processing of continuous speech. We applied this technique to the prosodic boundary detection on the output of a speech recognition application in order to test whether prosodic boundaries of different levels in tree confirm or not recognition hypotheses.
Two types of tree construction algorithms were tested: one using lexical information (word hypotheses), and another using only phonemic information (phoneme hypotheses). Both were successively used on the automatic alignment output ("perfect recognition" conditions) and on the ASR application output for the same spontaneous speech database so as to compare their applicability.