Language Model
Definition
A model that assigns probabilities to sequences of words, used to improve ASR accuracy by favoring linguistically plausible transcriptions.
A language model captures the statistical patterns of a language — which words tend to follow which other words, and which sequences are grammatically and semantically coherent. In speech recognition, language models help the decoder choose between acoustically similar alternatives.
Traditional ASR systems used n-gram language models trained on large text corpora. Modern end-to-end systems implicitly learn language modeling through their decoder component. External language models can still be used for rescoring or shallow fusion to improve accuracy on domain-specific vocabularies.