A new language model for speech recognition inspired by lin-guistic analysis is presented. The model develops hidden hierar-chical structure incrementally and uses it to extract meaningful information from the word history — thus enabling the use of extended distance dependencies —in an attempt to complement the locality of currently used trigrammodels. The structured lan-guage model, its probabilistic parameterization and performance in a two-pass speech recognizer are presented. Experiments on the SWITCHBOARD corpus show an improvement in both per-plexity and word error rate over conventional trigram models.