This invention relates to the generation of synthetic speech from conventional texts and in particular to the step in which a text in graphemes is converted into a text in phonemes. The grapheme text is analysed into rimes and onsets and each word is analysed from the end so that earlier occurring segments are at least partially defined by the identification of later occurring segments. It is a particular feature that an internal string of consonants, i.e. a string of consonants preceded and followed by a vowel, is split into two portions, namely a second portion which is contained in a database of onsets and an earlier portion which, together with the proceeding vowel or vowels, is contained in a database of rimes.