The language model describes to the ASR Decoder, the relationship between words and the probability of words appearing in a particular order, e.g., the language model will describe that the phrase “I went for a walk” is probable but “For a walk I went” is highly unlikely to occur. Language models may be domain specific, i.e., there is no need to include in the language model, the phrases a pilot would use if the end user is a doctor.
Speech recognition use either a statistical language model (SLM) or a context free grammar (CFG). A CFG commonly includes explicit definitions of the sentences that are expected to be spoken, whereas an SLM is a statistical representation of the probability of words being spoken in a specific order.
The Verbyx VRX implementation of a context free model includes many features that deliver speed and accuracy improvements over competing designs.
The choice of language model is dictated by the application. CFG provides the greatest level of accuracy and work well if the sentences that are likely to be spoken are well defined and constrained. SLM provides more flexibility in that the supported sentences are less constrained or explicitly defined but at the cost of poorer accuracy.