Language Modeling for Speech Recognition Frederick Jelinek Center for Language and Speech Processing Johns Hopkins University, Baltimore, MD 21218 A language model is a component of a speech recognizer whose task it is to assign an a priori probability to all conceivable utterances. Standard is a trigram language model which predicts the next word given the preceding two hypothesized words. Trigram language models are very potent, and it has become possible only very recently to improve on their performance. As of now, three types of language models exist: n-gram (which predict the next word from knowing the preceding n-1 words), structural (which base their prediction on a grammatical analysis of the preceding words), and neural (which carry out the prediction of next words with the help of artificial neural networks). The series of lectures will introduce the basic principles of all three language model types. It will then concern itself with the problem of automatic estimation from text data of the statistical parameters that determine model performance.