Abstract :
A speech recognizer is a device that translates speech into written text. As input, it takes the acoustic signal recorded by a microphone. As output, it produces a string of words intended to correspond to the input. The mapping from acoustic signal to a string of words is a complex task and it involves several stages. The central stage of this recognition process contains a search component, that makes use of two different sources of information, the acoustic model and the language model. Intuitively, the acoustic model gives a measure for how likely it is that a given chunk of the acoustic input corresponds to a given acoustic unit (e.g. phoneme, word). The language model gives a measure for how likely it is that this unit appears at this point in time, e.g. given the units that preceded it. Language models differ in the way the contexts are defined and in the way the probability distributions are being calculated. Different language models are presented and reviewed
