Start Free Trial
← All posts
From sound to text, live

How real-time speech recognition actually works

Why captions flicker and self-correct, why "upload your file" tools can't do live, and why 200 milliseconds is the line between reading along and reading behind.

4 latency stages 7 min read Published August 2026

Live captions look simple — sound goes in, words come out — but the machinery underneath is doing something genuinely hard: committing to words about a sentence the speaker hasn't finished. Understanding a little of how it works changes how you evaluate every tool that claims to do it. Two properties separate real live captioning from everything else: streaming architecture, which lets text appear while speech continues, and low latency, which decides whether the captions are usable in conversation or only afterward. No jargon required — here is the whole picture.

Streaming vs batch: two different machines

Most transcription tools you have met are batch systems: you hand them a finished recording, they take their time, and they return a polished transcript. Batch has a structural luxury — the model sees the entire audio before deciding anything. It can use what you said at minute 40 to fix a word from minute 2.

Streaming speech recognition gives that luxury up. Audio arrives in small chunks, and the model must emit words continuously, seconds — ideally milliseconds — behind the speaker, using only what it has heard so far. It is the difference between reviewing a chess game and playing one on the clock.

  • This is why "upload a file" tools can't do live. Their pipeline starts when the audio ends. No amount of speed turns a batch system into a live one — it is an architectural property, not a performance setting.
  • It is also why the two aren't scored the same way. A batch engine with full context should out-transcribe a streaming engine on the same audio; the streaming engine is solving a harder problem under a deadline.
  • And why "live" claims deserve scrutiny. Some tools fake liveness by transcribing short recorded chunks with a delay of several seconds. Real streaming shows words forming as the speaker talks.

Interim vs final results: why captions flicker

Watch any good live captioning tool closely and you will see words appear, then change. That flicker is not a bug — it is the engine being honest about uncertainty.

Streaming engines emit two kinds of output. Interim results are the model's best current guess, published immediately so you have something to read. Final results arrive a beat later, once enough context has accumulated for the model to commit — and they may revise the interim text. Say "I scream" and pause: a reasonable engine might show "ice cream" until your next words settle the question.

How to read the flicker when evaluating a tool: early guesses that get corrected quickly are the system working as designed. Final text that stays wrong is the actual error. When you test a captioning tool, judge the words that stop moving — that is the transcript you are really getting.

The latency budget: four places delay hides

"Latency" is the time from a word leaving the speaker's mouth to that word appearing on screen. It is a budget spent in four places, and every stage either earns or wastes milliseconds:

Capture

The tool has to acquire the audio: microphone buffers, or system/tab audio capture for meeting sound. Capturing system audio directly is both cleaner and faster than playing sound into a room and re-recording it through a mic.

Network

Audio chunks travel to the recognition engine and text travels back. Well-built tools stream over a persistent connection rather than posting files, and keep chunks small so no word waits long for a ride.

Inference

The model itself decides what was said. This is the hardest stage to make fast, because accuracy wants more context and more computation while latency wants less of both.

Render

The text has to reach your eyes — interface updates, smoothing, translation if you are running a second language side by side. Sloppy rendering can waste more time than the model spent thinking.

Why real-time feels simultaneous — and 1–2 s doesn't

Human conversation runs on tight timing: turn-taking, interruptions, and reactions all operate on fractions of a second. Captions that arrive within a couple hundred milliseconds land while the words still feel present — you read along with speech, the way you hear a person's voice and see their lips as one event.

Stretch the delay to one or two seconds and the experience changes category. The meeting has moved on; you are reading the previous sentence while deciding whether to respond to this one. For passive viewing — a lecture, a stream — that lag is tolerable. For anything conversational, it breaks the loop: by the time you have read enough to speak, your turn is gone. Interpreters feel this hardest of all, because their output is itself on a clock.

This is also why accuracy and latency trade off. An engine could always be a little more accurate by waiting longer — more context, better guesses. Live systems earn their keep by refusing to wait: publish a good interim now, finalize fast, correct honestly. The engineering achievement in a system that is both fast and accurate is holding both ends of that rope at once.

What this means when you choose a tool

  • Ask whether it truly streams. Words should form as the speaker talks. Multi-second chunks with a delay is batch wearing a live costume.
  • Judge final text, not flicker. Corrections are the design working; uncorrected finals are the errors that matter — our guide to word error rate shows how to count them properly.
  • Test latency conversationally. Read captions while someone talks to you and try to respond naturally. You will feel 200 ms versus 1,000 ms immediately, no stopwatch needed.
  • Feed it clean audio. Every stage of the pipeline does better with a good signal — the practical checklist is in how to improve live caption accuracy.
  • Compare on your real audio. Engines differ most on accents, terminology, and noise — the field guide in best live captioning tools of 2026 is where to start.

Frequently asked

How does live transcription work?

Live transcription streams audio to a speech-recognition model in small chunks, and the model emits words continuously while the person is still speaking — publishing fast interim guesses, then finalizing them as more context arrives. That streaming architecture is what separates live tools from batch tools, which need the complete recording before they start. Total delay comes from four stages: audio capture, network, model inference, and rendering.

Why do live captions change after they appear?

Because streaming engines publish interim results — their best immediate guess — and then revise them into final results once more context arrives. Hearing "I scream" versus "ice cream" may genuinely require the next few words. Quick corrections are the system working as designed; the errors that matter are in the text that stops moving. Judge any captioning tool on its final text.

What is good latency for live captions?

For conversational use — meetings, interpreting, phone calls — captions should land within a few hundred milliseconds of the words; Unicaption runs in real time, which reads as simultaneous. Delays of one to two seconds are tolerable for passive viewing like lectures or streams, but they break turn-taking: by the time you have read the caption, the conversation has moved on.

Why can't file transcription tools do real-time captions?

Because their architecture starts when the audio ends: batch tools process a complete recording, using the whole file as context. Real-time tools are built differently — they accept a continuous stream and commit to words with only partial context, under a latency deadline. Speed doesn't convert one into the other; a tool either streams or it doesn't.

See streaming done properly

state-of-the-art streaming accuracy in real time, with live translation in 60+ languages. Watch it keep up with real speech — free 30 minutes every week.

Start Free Trial →