This thesis presents an approach to create topic dependent language models. It is shown that a gain of 5% was reached using speech recognition for news broadcast. First, this document presents the theory, on which language models are based and the problem of sparse data which is one of the biggest problems associated with the creation of adequate language models. Next, general guidelines are presented in regard of the choices and techniques implied in the creation and the adaptation of language models. Finally, experimental results are presented and commented on to sustain the concept of topic dependent language models. |