Glossary
A language model (LM) assigns probabilities to word sequences. In speech recognition, an LM disambiguates between phonetically-similar transcripts: "recognize speech" vs "wreck a nice beach" sound nearly identical, but the LM gives the first one much higher probability based on word co-occurrence patterns. Modern transformer ASR systems like Whisper integrate the LM directly into their decoder.
Old-school n-gram LMs
Pre-2018 ASR used n-gram language models — count-based statistics over training text, smoothed to handle unseen sequences. Typical: 3-gram or 4-gram LM trained on billions of words of news and web text. Decoding combined the acoustic model score and LM score, weighted by an LM scaling factor. Effective but capped — n-gram LMs can't reason about long-range context.
Neural LMs in modern ASR
Whisper's decoder is itself a transformer language model conditioned on encoder output (audio). It outputs one token at a time, each conditioned on all previous tokens — effectively an unbounded-context LM. This is why Whisper handles long sentences with proper noun consistency and grammatical agreement that older systems missed. The downside: hallucinations on long silences, where the LM happily generates plausible-but-wrong text.
Domain adaptation via LM
If you transcribe medical, legal, or technical content with lots of jargon, the general-purpose LM may misrecognize specialized terms. Solutions: (1) biasing — provide the model a hint list of likely terms (Whisper supports this via the "prompt" parameter); (2) fine-tuning — train on your domain's transcripts. We use Whisper's prompt parameter for known-domain uploads and otherwise let the general LM handle it.
In practice
Audio: "the patient presented with paroxysmal atrial fibrillation." A general-purpose LM might transcribe "paroxysmal" as "box of small" — phonetically close, semantically nonsense. With a medical-vocabulary prompt, Whisper biases toward the correct medical term and gets it right.
Related terms
Further reading
Try AI transcription
Get started — freeLast updated: September 20, 2026
Frequently Asked Questions
What does Language model mean in transcription?
The component that turns phoneme guesses into actual words by predicting likely word sequences. Modern Whisper-style transformers integrate both into one model.
Why does Language model matter when choosing a transcription workflow?
Language model affects how you understand transcript quality, timing, compatibility, or the technology behind speech-to-text results.
Where can I apply Language model on Transcript.you?
Use the related workflow at /ai-transcription when you want to see how this glossary concept connects to an actual transcription task.
Browse the full transcription glossary or read the complete guide to AI transcription.