Purpose:
The description of the distribution and estimated probabilities of words. In speech recognition systems, `stochastic language models' are used to provide top-down hypotheses for limiting the number of hypotheses generated by the speech recognition decoding component. This is done by calculating the probability of a word (or other unit), given the n preceding words (n-gram model).
The formalism most frequently used to construct these models is the Hidden Markov Model formalism. A Hidden Markov Model (HMM) is a probabilistic finite state device in which transitions and states are assigned probabilities. These probabilities are estimated from an examination of the distribution of the words concerned in a corpus.
The principle of using HMMs is to find the string of words with the maximum probability, given the input from the signal decoder.