A language model predicts how likely a sequence of words is. In speech recognition it complements the acoustic model, biasing output toward wording that makes sense so the system prefers "recognize speech" over "wreck a nice beach." Where the acoustic model hears sounds, the language model encodes what people actually say.
How a language model works
A language model assigns a probability to a string of words, usually by predicting each word from the words before it.
Classical systems used an n-gram model, which estimates the odds of a word from the previous one or two, counted across a large text corpus. Neural language models replaced fixed n-gram counts with a network that learns longer-range context.
Either way the recognizer queries the model during decoding and multiplies its judgment against the acoustic score, so two candidate transcripts that sound alike are broken apart by which one reads as more natural language.
Classical ASR versus modern ASR
Traditional systems trained a standalone language model and combined three scores at decode time: the acoustic fit to the audio, a pronunciation lexicon mapping words to sounds, and the language model's plausibility judgment. That trio is how the system chose among homophones.
End-to-end neural models learn an implicit language model straight from transcribed audio, so a separate component is no longer required to produce a transcript.
External language models still get layered back in through fusion when a system needs domain adaptation, and a large language model is increasingly used as a post-processing pass to punctuate, clean, or rescore raw output. The principle has not changed: give the recognizer a sense of likely wording to sharpen its guesses.
Why the language model matters
The language model is where most provider customization lives. Features like a custom language model, word boosting, and custom vocabulary all work by nudging the language-model side toward names, product terms, and jargon it would otherwise miss.
If your audio is full of unusual proper nouns, that customization surface matters far more than a fraction of a point of benchmark accuracy. Model quality here is often measured with perplexity, a score of how well the model predicts held-out text.
Frequently asked questions
What is a language model? It is a system that estimates how probable a sequence of words is. In speech recognition it ranks candidate transcripts by how natural they read, resolving words that sound identical.
What is the difference between a language model and an acoustic model? The acoustic model decides which sounds are in the audio, while the language model decides which word sequence those sounds most likely form. Classical decoders combined both scores.
How do you make a speech recognizer learn new words? Through the language model side, using custom vocabulary, word boosting, or a custom language model trained on your domain text, so names and jargon become likely instead of unexpected.