An acoustic model can determine the probability of phonemes given a piece of audio, but this alone is insufficient for accurate transcription. For example, the phrases "recognize speech" and "wreck a nice beach" can sound very similar. An acoustic model might assign high probabilities to both interpretations. To resolve this ambiguity, the system needs to understand which sequence of words is more likely in the given language.

This chapter introduces the language model, the component responsible for adding linguistic context to the recognition process. By assigning a probability to a sequence of words, the language model helps the ASR system choose the most plausible transcription from a set of acoustically similar candidates.

You will learn about the following topics:

• The function of a language model in calculating the probability of a word sequence, or $P(\text{word sequence})$.
• Classic N-gram models, which estimate the probability of a word given the preceding N-1 words, such as a bigram model that uses $P(w_n | w_{n-1})$.
• How probabilities are calculated from large text corpora.
• Perplexity, a standard metric for evaluating how well a language model predicts text.
• A brief look at how neural networks are used to build more advanced language models.