Evaluation of a segmentation system based on multi-level lattices

This paper addresses the problem of the evaluation of an automatic continuous speech segmentation system based on multi-level lattices called dendrograms previously described in [2] [3]. After a short overview of the segmentation framework, we briefly discuss the difficulties inherent to such an evaluation. We first state quantitative results based on the usual quality coefficient and oversegmentation rate measures obtained for the automatic paths selection yielded by our system and we estimate the potential improvement margin of our approach. These quantitative results are completed by a qualitative analysis of the system performance by examining the detection accuracy of all possible transitions between 6 broad phonetic classes. The results are judged encouraging given the restricted number of phonetic likelihood criteria used at this stage of development and demonstrates the feasibility of the segmentation process.