Language Model

The language model is the component of a speech recognizer which decides whether a particular sequence of words is a likely utterance in some human language. In the so-called FundamentalEquation of speech recognition:

$\begin{displaymath} \hat S = \arg\max_S P(X|S,\lambda) P(S|\lambda) \end{displaymath}$

The language model is the source of the PriorProbability $P(S|\lambda)$ , that is, the language model tells us what the probability of a sentence S is, given the model $\lambda$ .

In most medium to large vocabulary speech recognition systems, a so-called N-Gram language model is used. This type of model is based on the fact that, given the ChainRuleOfProbability, the probability of a sentence can be decomposed into the probability of each word given all of the previous words, or the history. The N-Gram model is based on the assumption that only the last N-1 words are actually relevant in predicting the current word. As linguists we know this to be completely false, but in practice it works very well.

Why is this? Primarily, the N-Gram model is very easy to train from large amounts of unannotated or minimally unannotated text.

The CMU Sphinx Group Open Source Speech Recognition Engines

Language Model