Minimal Rules for Synthesizing Speech

It has been found to be extremely difficult to isolate phonemic elements from recorded utterances or to synthesize speech by assembling prerecorded phonemic segments. One reason for the difficulty lies in the fact that the perceptually discrete phonemes are typically combined, and in some cases encoded, into units of essentially syllabic dimensions. As a result, prerecorded elements must, in many cases, approximate syllables, and the synthesis of speech by this means will require a large inventory of recordings. By taking advantage of knowledge about the acoustic cues for speech perception, however, it is possible to write rules for synthesis in terms of phonemes (rather than syllables) and thus reduce considerably the number of separate rules or items needed. Indeed, one can reduce the number of rules still further by writing them at the level of subphonemic dimensions, viz., place and manner of articulation. Several complicating factors make it impossible to achieve an ideal minimum. First, rules or rule modifiers must be added to take care of certain prosodic and positional variations. Failure to do so not only affects naturalness, but also impairs intelligibility, even at the level of segmental phonemes. Second, it is necessary in a few special cases to have different rules for a single consonant phoneme (or a dimension of that phoneme) before different vowels. This reflects the occasionally complex relation between phoneme and articulation on the one hand and sound output on the other; presumably, this complication would not affect the rules of synthesis for an articulatory model. A system of rules for synthesis framed largely in terms of subphonemic dimensions is described with reference to an example. Words and sentences of rather high intelligibility have been synthesized by such rules.