The work reported here represents an attempt to extend the collaborati
ve view of conversation (see, for example, Clark and Wilkes-Gibbs, 198
6) to the level of the characteristics of individual words - in partic
ular, word duration. Speaker dyads engaged in the tangram task which h
as been widely used for researching discourse-level behaviour. Four mo
nologue and four dialogue trials were recorded from each speaker. Cont
ent words which appeared in at least seven trials were identified, and
the successive tokens of these words were digitised and measured. Ove
rall, words spoken in monologues were reliably longer than words spoke
n in dialogues, suggesting that speakers were conservative in their es
timate of how intelligible a listener would find the words when no vis
ual or linguistic feedback was available. In addition, monologue and d
ialogue tokens exhibited different shortening effects in response to r
epetition; in dialogue, speakers attenuated word durations on the seco
nd mention, but in monologue no shortening occurred between successive
repetitions, although a gradual decrease in duration was observed acr
oss the four trials. These results parallel findings relating to refer
ring expressions, where the decline in the number of words used to ref
er to an entity is much sharper in dialogue than in monologue. The res
ults are discussed within the framework of the collaborative model.