Speech recognition and models
Language model
A model that predicts likely word sequences, used both in transcription and to write summaries.
A language model is a model trained on large amounts of text, which predicts which words are likely to follow one another.
In transcription it plays two roles.
1. It settles the ambiguous cases
Audio alone is often ambiguous. “We will take it in the review meeting” and “we will take it in the review, meeting” sound identical. The model picks the word sequence that is most likely in context.
This is also the explanation for a pattern many people recognise: the system does well on ordinary sentences and misses proper nouns. Ordinary word sequences the model has seen millions of times. Your surname it has never seen.
2. It writes the summary
A transcript from an hour-long meeting is around 8,000 words. Turning that into a summary with decisions and tasks is a job for a language model, not for the speech recognition. They are two entirely different links in the chain, even though they are often sold as one product.
Large language models
LLM, large language model, is used for the big models of the GPT and Claude type. They are language models in the same sense, only much larger and more general. In a transcription context they are almost always used in step 2, for summarising and structuring, not for the speech recognition itself.