Summary The “cloze” procedure, which requires raters to fill in words deleted from a transcript and provides a measure of the predictability of speech, was used to study the predictability of the various grammatical classes and of specific items drawn from them, and changes in such predictability with systematic variation in rate of textual deletion (from every second item deleted to every sixth item deleted). A transcript of running speech elicited by TAT cards constituted the text and two measures were obtained indicating success in verbatim identification (V) and success in form-class identification (FC); the relation between these two measures was also examined. The main findings are (a) that success both in FC and V completion increases moderately with decreasing frequency of deletion, but that raters do fairly well, particularly for FC scores, even when every second item is deleted; (b) that there are considerable differences in performance between grammatical classes both for FC and V scores, these differences being less for the former scores than the latter; and (c) that within any given grammatical class there are still substantial differences among the items with regard to both verbatim and form-class predictability. It was argued that the determinants of FC and V predictability are rather different, FC predictability being more dependent upon the relatively close grammatical environment whereas V predictability depends more on both close and remote topical content or semantic features of discourse. It was also suggested that these determinants may vary with the particular grammatical class under consideration. While grammatical or form-class contraints are tighter than contraints on specific identification it was suggested that the V or V/FC measure provides an underestimate of the actual predictability of the sense of the discourse, and that the communicative significance of a particular level of V or V/FC success may vary with the grammatical class involved, being greater for semantic than for syntactic items.