In short
A transcript records speech. A subtitle is a readable event on a picture. Both can live in an
.srtor a.vtt. That is why the two get sold as the same product.They are not. A transcript follows the speaker. A subtitle follows the viewer. If you need timed text for a series, a film, or a linear slot, you need the second object.
This piece stays with that distinction. Captioning for deaf and hard-of-hearing viewers, and translation choices inside a subtitle, are separate jobs.
What speech-to-text actually returns.
Automatic speech recognition (ASR) writes words. It can add speaker labels. It can add word or segment times.
The default path is a transcript. Public speech-to-text documentation treats word or segment timestamps as an extra request, described as data for transcripts and video edits. The same guides treat subtitle file formats as a specialised output, not as the default transcript.
Those timestamps follow how fast someone talked. They do not carry a reading budget, a two-line layout, or a gap between events. Those are subtitling decisions. They happen after the transcript exists.
The file can say "subtitle" and still be a timed transcript.
Speech-to-text wrappers often add a subtitle-shaped export. You upload audio or video, get text, then download SubRip (SRT), WebVTT (VTT), or burned-in video. The label on the button is "subtitles." The constraints that make a broadcast subtitle are not in the speech-to-text step.
An .srt or .vtt is a timed text file: numbered events with a start time, an end time, and some words. That is all the format promises. It does not promise that a viewer can read those words on a picture, at a workable speed, without covering the shot. The World Wide Web Consortium (W3C) publishes the W3C WebVTT specification. That specification requires end to be later than start, and it lists cues by start time. It does not set characters per second (CPS), characters per line (CPL), minimum duration, or shot-change rules. It even allows different cues to overlap. A player can open that file. You still do not have a subtitle a viewer can read.
Undertitle makes professionally timed subtitles for broadcasters, streamers, distributors, and localization teams. Editors keep the final call. That is a different product category from a transcript export.
A live transcript is not a subtitle.
Live speech-to-text can print words as they arrive. A subtitle cannot be checked for line length or reading speed until the chunk of text is known.
So you need delay. Keep the words long enough to break lines, test duration, and leave a gap. Then put them on screen.
Key point
Live AI transcript tools wrap speech-to-text models. They do not write a subtitle. The model underneath is the product, from houses such as Meta (Muse Voice Transcribe), OpenAI, Google Cloud Speech-to-Text, and others. A wrapper can stream words and export a subtitle file. That file is still a transcript with times.
Live subtitling already works with delay. Respeaking and live encoding insert time between the spoken word and the on-screen event, because the word has to be heard, turned into text, and formatted. Moores et al. (JoSTrans, 2020) treat that latency as inherent, not optional. A zero-delay word stream skips the formatting step. It is a live transcript.
Transcript vs subtitle.
A transcript covers a short list of jobs. A subtitle has to cover more.
| Job | Transcript | Subtitle |
|---|---|---|
| A written record of what was said | Yes | No. The on-screen text is condensed and reshaped. |
| Words while a live event is still happening | Yes | No. The event has to wait for line, duration, and gap checks. |
| Readable lines on a picture (length, two-line layout) | No | Yes |
| Duration and reading speed a viewer can keep up with | No | Yes |
| Gaps, no overlap, timing around cuts | No | Yes |
| A file that opens in a player | Yes, if it has times | Yes. That still does not make it a subtitle. |
Parameters a real subtitle has to meet.
You set the exact numbers. If a file never encoded these checks, quality control (QC) will. The figures below are typical practice from professional and research sources. They are not the published limits of named broadcasters.
| Parameter | What it means for the file |
|---|---|
| Time order | The event must exist in time. End before start is a broken cue. WebVTT requires the end timestamp to be later than the start. |
| Text present | A subtitle is written language on screen. An empty event is a hole in the file. SUBTLE (2023) defines subtitles as lines of written text. |
| Line count | More lines cover more picture. Usual maximum is two lines at a time (Karamitroglou, 1998). SUBTLE: 1 or 2 lines; 3 only in rare cases. |
| Line length | A line that is too long shrinks the type or runs into the frame. Typical range is about 34 to 50 characters including spaces (SUBTLE). Streaming practice has moved from about 32 toward 42 (Szarkowska and Gerber-Morón, 2018). |
| Maximum duration | Leave a subtitle up too long and fast readers re-read it. Typical cap is about 6 seconds for a full two-liner, about 3.5 for a full one-liner (Karamitroglou, 1998). SUBTLE matches the 6-second ceiling, with limited exceptions such as songs. |
| Minimum duration | A flash is not readable. SUBTLE: avoid under 1 second where possible. Karamitroglou: a single-word subtitle at least 1.5 seconds. |
| Reading speed | The viewer has to read the event and watch the picture. Keep average speed at 12 to 15 CPS, or 150 to 180 words per minute (WPM); do not let peaks exceed 16 to 17 CPS (SUBTLE). Some current streaming files run at 17 to 20 CPS (Szarkowska and Gerber-Morón, 2018). |
| Dialogue formatting | Two speakers in one event have to read as two voices. Karamitroglou: a dash plus a space at the start of each dialogue line, left-aligned. SUBTLE: follow the national punctuation and dialogue-indicator rules for the delivery. |
| Overlap | Two events in the same frames fight for the same reading slot. WebVTT allows overlapping cues; production QC does not. EZTitles flags overlapping in/out times; SUBTLE requires a gap instead. |
| Gap between cues | With no pause, two similar events look like one subtitle. Karamitroglou: about 1/4 second between consecutive subtitles. SUBTLE: a fixed interval of 2 to 6 frames (3 to 4 recommended). |
| Duplicate consecutive text | The same words in the next event are not new information. Treat adjacent identical text as a defect. It is a standard structural test in subtitle validators (see timecode.dev). |
| Rapid cue bursts | A run of tiny events is unreadable even when each timestamp is valid. Combine minimum duration and gap. SUBTLE: avoid long chains of short subtitles. |
| Short single-line runs | Isolated one-liners in a row waste the two-line budget. SUBTLE: each event should be semantically self-contained. Merge or re-segment sentences that span three or more events. |
| Shot change | A subtitle that sits across a cut looks late, early, or stuck to the wrong picture. Time to the dialogue and avoid crossing shot changes when you can (SUBTLE). Drop the subtitle before a cut that changes the theme; ordinary camera-size changes need not (Karamitroglou, 1998). |
None of this is in a raw transcript. You either build it, or you hand the file to an editor who will.
Certainly ask for this.
Ask what the file was measured against. "It opens in the player" is a format check. It is not a subtitle check.
A quality report that scores the parameters above is the useful next artefact. Treat an SRT from speech-to-text as a transcript with times, unless someone can show the reading, line, duration, gap, and shot-change work already happened.
If those checks are missing, you still have a timed transcript. If you need a subtitle that already meets them, book a free demo with Undertitle. Editors keep the final call.
Frequently asked questions
01
Is an SRT from speech-to-text a subtitle?
No. An .srt or .vtt stores timed events: a start, an end, and some words. That is all the format promises. The file format is not the product. Timed text in an SRT is still a transcript until reading, line, duration, and gap checks have been applied.
02
Why is a live transcript not a subtitle?
A subtitle needs the full chunk of text before you can break lines, test duration, and leave a gap. That delay is the job. A live word stream does not wait for it. The result is words on a timeline, not a readable on-screen event.
03
What should I ask before I accept a subtitle file?
Ask for the quality report, not only the file. Request the reading speed, line, duration, gap, overlap, and shot-change results. "It opens in the player" is a format check. If those checks are missing, you still have a timed transcript.
04
What is the difference between a transcript and a subtitle?
A transcript is a written record of what was said. It can carry times, and it can print words while a live event is still happening. A subtitle is not that record. The on-screen text is condensed and reshaped so a viewer can read it: line length, a two-line layout, duration, reading speed, gaps, no overlap, and timing around cuts. A file that opens in a player is not enough. A timed transcript can open too. That still does not make it a subtitle.
Book a free demo
Want to see how Undertitle turns English shows and films into Dutch subtitles? We will walk through it with you.