Abstract: If we can model the cognitive and communicative processes underlying speech, we should be able to better predict what speakers will do, and thus improve language models. This paper presents an initial exploration of this idea. In the Switchboard corpus, we find that word probabilities vary with various non-lexical indicators of cognitive and communicative states, including local volume, local speaking rate and other prosodic features, and also time since start of utterance and since since other reference events. Conditioning word probabilities on 8 such features improved word predictions, reducing the perplexity by 4.4% relative to a trigram baseline.