Guide

Real-time transcription: how live speech becomes text in under a second

By the exma team · August 5, 2026 · 8 min read

What is real-time transcription?

Real-time transcription — also called live transcription or streaming speech-to-text — converts speech into text while it is being spoken. Batch transcription reads a finished recording with full context and returns a transcript in minutes; a streaming engine gets no second listen. Audio arrives in small chunks, text must appear almost immediately, and the transcript assembles itself in front of you.

The distinction matters because the two modes serve different jobs. Batch is for turning yesterday's recording into a document. Real-time is for the moments where the text has to exist now: a caption on a screen, a read-back in a hearing, a searchable draft of a meeting still in progress.

How live transcription stays fast and accurate

The trick that makes streaming work is interim and final results. The engine shows a provisional hypothesis within a fraction of a second, then revises it as more acoustic evidence and sentence context arrive, and finally commits once the phrase is settled. Watch a live transcript closely and you can see the last few words shimmer and self-correct as the sentence completes — that isn't a glitch, it's the visible half of the latency-accuracy trade.

The rest of the pipeline runs alongside recognition: speaker diarization separates voices and labels turns as they happen, punctuation and casing are applied on the fly, and timestamps keep every word tied to its position in the audio. For a deep walkthrough of that pipeline in the hardest setting we know, see real-time speech-to-text in the courtroom.

What "real time" actually means, in numbers

• ~100–500 ms — first provisional words appear. This is what makes captions readable against the speaker's lips.
• 1–3 seconds — text finalizes, once the engine has enough context to commit.
• Beyond a few seconds — the transcript stops feeling live. Read-backs lag the room, captions trail the conversation, and the draft becomes a delayed recording instead of a working tool.

Latency comes from more than the model: audio chunking, network round-trips, and how conservatively the engine waits before finalizing all add up. Good tools tune the whole path, not just the recognition step.

Where real-time transcription is used

Accessibility and captioning

Live captions let deaf and hard-of-hearing participants follow lectures, broadcasts, and meetings as they happen. This is the oldest and largest use of streaming speech-to-text, and the reason sub-second latency became table stakes.

Meetings and calls

A live transcript turns a meeting into a searchable document while it is still running — no "what did she say ten minutes ago," just search. Consumer note-takers layer summaries on top; see how they compare to a verbatim tool in exma vs. consumer AI note-takers.

Legal proceedings

Depositions, hearings, and courtrooms use live drafts for instant read-backs and same-day prep, with human review turning the draft into the certified record afterward. This is exma's home turf: domain-tuned vocabulary, speaker labels, and capture that never joins the proceeding as a participant.

Healthcare, education, broadcasting

Clinical scribing during consultations, live lecture notes, and broadcast subtitling all run on the same streaming machinery, each with its own vocabulary and privacy constraints.

Choosing a real-time transcription tool

• Latency you can feel. Demo it live. If the text lags the speaker enough to notice, it will lag every meeting.
• Accuracy on your audio. Test with your accents, your jargon, your meeting rooms — not the vendor's demo clip.
• Speaker labels in real time. A wall of unattributed text is a caption, not a transcript.
• Capture method. Bot participants only work on platforms they integrate with, and are visible to everyone in the call. Source capture — microphone plus system audio — works everywhere and stays out of the room. exma captures at the source; there is no bot to admit.
• What happens afterward. Live text is half the job. Look for synced timestamps, editing, exports (TXT, DOCX, PDF), and — for regulated work — a path to human review and certification. Our security checklist covers the data-handling questions.

Try it in ten seconds: the exma.ai homepage demo streams your microphone through the full pipeline — speak and watch provisional text appear, shimmer, and lock in. No account needed.

Frequently asked questions

How accurate is real-time transcription?

Close to batch accuracy on the same audio — well above 90% in reasonable conditions, higher on clean speech. Audio quality (microphone placement, noise, crosstalk) moves accuracy more than engine choice does.

What latency counts as real time?

Provisional words within ~100–500 ms; final text within one to three seconds. Slower than that stops feeling live.

Does it work for video calls and phone calls?

Yes — if the tool captures system audio digitally, it transcribes any platform without joining as a bot. Tools built on bot participants are limited to the platforms they integrate with.

Can a real-time transcript be the official record?

The live stream is a working draft. Review against the audio and certification make it official — the full requirements are in what makes a transcript court-admissible.