Large language model (LLM)

A large language model (LLM) is a language model with a very large number of parameters. Parameters are the numbers that training adjusts. Google's Machine Learning Glossary says that, at a minimum, an LLM is a language model having a very high number of parameters ¹. More informally, the term covers any Transformer-based language model, and the glossary names Gemini and GPT as examples. The systems behind most modern chatbots and writing assistants are LLMs.

Scale is the defining feature

The 2020 GPT-3 paper described a model that was "an autoregressive language model with 175 billion parameters, 10x more than any previous non-sparse language model" ². The paper reports that scaling up the model greatly improved task-agnostic, few-shot performance, meaning the model handled new tasks from a few examples in the prompt, with no gradient updates for each task.

How it works: predicting the next token

An LLM reads text as tokens, which are pieces of words, and predicts the next token from the ones before it. Each prediction is appended to the text and the process repeats, which is why the model is called autoregressive. Repeated many times, this produces a sentence, an email or a page. Training on very large amounts of text makes the model good at grammar, at the style of its input and at many facts in its training data.

A concrete example: a rewrite of a dictated sentence

Give an LLM the transcript "the report is late because uh the data came in late" and ask it to make the sentence clearer. It may return "The report is late because the data arrived late." The model is predicting a likely continuous rewrite, not checking the words against your audio. The new sentence reads well, but it is only as faithful as the model's guess about what you meant.

Why it matters for dictation

LLMs can correct speech recognizer output. A 2023 study on the HyPoradise benchmark found that an LLM can correct words missing from the recognizer's N-best list, because it generates text instead of choosing among candidates ³. That is useful when the right word was never offered.

The same ability causes trouble. A study of LLM correction for speech recognition reports that LLMs make many unnecessary changes to the input transcription, which the authors call over-correction ⁴. A correct word may be replaced by a more common one, so a dictated name or technical term can change silently. Check the output before you send it. Our guide to speech to text and LLMs explains where the model fits in a dictation pipeline.

Related terms

• Generative AI: the wider class of models that create content, including LLMs.
• Speech language model: an LLM that takes audio as input instead of only text.
• Generative error correction (GER): using an LLM to rewrite speech recognizer output.
• Prompt (AI): the input you give an LLM to get a response.
• Language model (in speech recognition): the older statistical version of the same idea, used to choose likely words.

Guides that cover it

• Speech to text and LLMs: what language models change
• Speech to text for AI: why spoken prompts work better
• AI dictation: how AI voice typing works
• Speech to text AI: dictation, assistants and voice agents