The system also provides a kind of primitive phonological framework, in that it allows the statement of ordered, contrast-dependent allophone rules which modify the stored data for synthesis of a phone. The description of the contexts in which a rule can be altered are built up from a limited set of binary contextual features, e.g. 'prevocalic', 'post-vocalic', 'stressed': these contextual features are part of the program. Thus it is claimed that the nature of phonological capacity is such that only a small fraction of the conceivable contexts in fact occur in the phonological rules of natural languages -- a claim with which Chomsky and Halle (1968: 400-1) appear to be sympathetic. The program has been used to synthesize both the General American and the Southern British dialects of English (Haggard and Mattingly 1968).

In this system the prosodic rules are phonetic: phonologically predictable stress and intonation effects must be marked in the input. Vanderslice (1968), however, has proposed a set of rules for predicting the occurrence, in English, of the prosodic features for which his definitions have been given above, in particular accent. His strategy is to assign provisional accents to all lexically- stressed syllables and then to delete certain of these accents. For example his 'rhythm rule' deletes the middle one of three consecutive accentable syllables in the same sense group. If a word such as 'unknown' is assumed to have two accentable syllables in its lexical form, this rule accounts nicely for the shifting stress in such words. Other rules proposed by Vanderslice rely on syntactic or semantic conditions: these conditions will somehow have to be marked at the input to the phonology.

Finally, mention should be made of Allen's (1968) programming of the Chomsky and Halle (1968) rules for the assignment of accent. At this writing, so far as we know, no one has thus far attempted a program for synthesis of English based on the Chomsky-Halle rules for segmental phonology, or even a computer simulation with phonetic-feature matrices as output. Fromkin and Rice (1970), however, have developed a program for which the input format follows closely that of the Chomsky-Halle phonological conventions, and permits the testing of a set of phonological rules.
 

We must now try to sum up the current state of synthesis by rule and to indicate the directions the work may be expected to take in the future. As we have seen, it is possible to synthesize speech by rule which is not only intelligible but also reasonably acceptable to a native speaker. Moreover, the trend of research in the past few years has been toward the development of systems of increasing phonetic sophistication with correspondingly greater theoretical interest. In the synthesis of segmental sounds, the emphasis has shifted from acoustic models to vocal-tract shape models, and from shape models to articulator models; a similar trend is