Spoken language speech recognition systems need better understanding of natural spoken language phenomenon than their dictation counterparts. Current language models are mostly based on written text and/or very tedious Wizard of Oz or real dialog experiments1. In this paper we propose to use Internet documents as a very rich source of information for spoken language modeling. Through detailed experiments we show how using Internet we could automatically prepare language models adapted to a given task. For a given recognition system using this approach the word accuracy is up to 15% better than a system using language models trained on written text.