We feel our model is adequate as a phrase break assignment algorithm as it is simple, performs well and does not have unrealistic input requirements. Both the model and the POS tagger have been fully implemented and are used as the standard phrase break assignment model in the distributed version of the Festival Speech Synthesis System [2].
Although we feel we have thoroughly tested this model and investigated the varying of its basic parameters to find the (near) optimum values, we feel that there is ultimate limit to how well such a model can perform with only part of speech information. It has been shown that some decisions about phrase break assignment can only reasonably be made with additional syntactic information. For example, identification of the use of prepositions as verb particles is an identifiable mistake in our (and other's) algorithms. Also the verb balancing rule as described in [1] is a valid phenomena which our current algorithm cannot capture without the addition of some form of syntactic information. There is of course a trade-off between time to run a reliable parser and the possible improvement in results, but we still feel such investigation is worthwhile, if only to find out the benefits it would give in improving on our current results.