What is Real-Time Transcription?
Also called streaming transcription, live transcription, live captions.
Real-time transcription converts speech to text continuously while a person is still speaking, emitting provisional results that are revised as more audio arrives. It differs from batch transcription, which processes a complete recording after the fact and can use the whole file as context. The live variant trades some accuracy for immediacy.
A streaming recognizer consumes small audio chunks and emits interim hypotheses within a few hundred milliseconds. Those hypotheses are unstable by design: as later audio disambiguates earlier sounds, words are rewritten. Systems mark results as partial until a stability point or an endpoint is reached, at which stage the text is declared final and downstream consumers can safely act on it.
Accuracy is structurally lower than batch mode because the model sees limited right-hand context and cannot revisit the whole recording. Longer lookahead windows raise accuracy and latency together. Punctuation, casing, and number formatting are usually applied by a separate post-processing stage that also runs incrementally, which is why live captions sometimes gain commas a beat after the words appear.
Live captions are the most visible application, and their quality is judged on more than word error rate. Caption presentation rules cover reading speed, line length, how many lines stay on screen, and whether corrections flicker distractingly. Broadcast and accessibility guidance treats latency and stability as first-class metrics alongside correctness.
For agents, streaming transcription is what allows a reply to be planned before the speaker finishes. Partial transcripts feed endpointing models, intent detection can start early, and retrieval can be kicked off speculatively. The tradeoff is that any work started from a partial hypothesis may need discarding when the text is revised.
Key points
- Emits revisable partial results while speech continues
- Lower accuracy than batch because right-hand context is limited
- Punctuation and formatting are applied incrementally
- Caption quality depends on stability and reading speed, not just accuracy
- Partial transcripts let agents start work before the turn ends
In practice
During a webinar, a caption bar shows text roughly a second behind the speaker. The word 'their' appears, then rewrites to 'there' once the following clause arrives, then gains a comma from the punctuation stage. The same stream feeds an assistant that has already begun retrieving a pricing document from the partial transcript, and it discards that lookup when the corrected text changes the topic.