Automatic subtitles take about a minute to generate and about ten minutes to get right. A speech recognition engine listens to your audio and writes timed text, and you then fix names, punctuation, line breaks and timing before you export. This guide covers the whole path: how the technology works, a repeatable workflow, which file formats to use, how to translate subtitles, and where AI dubbing, voiceover and text to speech fit when subtitles are not enough.
TL;DR
- Auto captioning turns speech into timed text with speech recognition. It is a fast first draft, never a finished product, so always review it.
- Check proper names, numbers, speaker changes, sound cues and timing before you export.
- Use a sidecar file (SRT or WebVTT) when the platform supports it, and burned-in captions for short vertical clips where you control the look.
- Translated subtitles must respect reading speed. Dubbing needs a different script, adapted to mouth movement.
- Pick tools that charge for what you use, so a single clip does not lock you into a monthly plan.
Table of Contents
- How automatic subtitles actually work
- A workflow that works for most videos
- Subtitles, captions and non-speech sound
- Burned-in captions or a subtitle file?
- Subtitles for Reels, Shorts and TikTok
- Translating subtitles with AI
- AI dubbing, voiceover and text to speech
- What to review before you export
- Our perspective: captions are part of the edit, not an afterthought
- Where Brix36 fits
- FAQ
- Sources
How automatic subtitles actually work
Auto captioning relies on automatic speech recognition (ASR), which converts spoken words into on-screen text. As one university accessibility guide explains, ASR software guesses what someone is saying from voice patterns, and natural language understanding (NLU) engines then use the surrounding words to infer meaning. That second step is why a good engine can tell "their" from "there" in a clean sentence.
The output is more than a block of text. Speech services return the offset and duration of each recognized phrase, and for live use they stream partial results before delivering a complete final transcription once recognition finishes. Those offsets are what make the words appear in sync with the picture.
The same guide warns about the weak spots: captions can misinterpret words, context or sentence structure, and they can drift out of sync. Treat the machine as a fast typist who has never heard your topic before.
What trips the engine up
You will see the same problems in almost every draft:
- Names and jargon. Artist names, product names, place names and technical terms are rarely in the engine's vocabulary.
- Music under speech. A loud bed makes consonants hard to hear, and sung lyrics are often mis-transcribed.
- Overlapping voices. Two people talking at once usually produce one garbled line.
- Accents and fast delivery. Accuracy drops with heavy background noise, distance from the microphone and rapid speech.
- Numbers and dates. "Twenty-fifteen" may become "2015" or "2050" depending on context.
The cheapest fix is upstream: record clean audio, keep the voice louder than the music, and say names clearly once.
A workflow that works for most videos
The order below keeps you from fixing the same line twice.
- Lock the edit first. Generate subtitles only after your picture is cut. Every trim after that shifts the timing and forces a re-sync.
- Export a clean audio track or upload the finished video. Dialogue-only stems give better recognition than a full mix with music.
- Choose the spoken language correctly. Many tools detect it automatically, but a wrong guess on a mixed-language clip ruins the whole transcript.
- Generate the draft. Expect something between a rough and a very good first pass, depending on audio quality.
- Read the transcript without the video. Text errors are easier to spot when you are not watching the picture. Fix names and numbers first.
- Watch with sound on, then with sound off. The first pass checks sync; the second checks whether the video still makes sense to a viewer on a train.
- Style and position. Choose a readable font, strong contrast and a safe area so captions do not collide with interface elements or lower thirds. Our guide to lower thirds animation explains how to keep a safe margin so titles and captions do not fight for the same space.
- Export in the right form (see below) and test it on the platform where it will live.
Line breaks and pacing
Most editors aim for two lines at most, broken at natural phrase boundaries rather than mid-clause. Split a long sentence into several cues instead of cramming it into one. Keep each cue on screen long enough to read comfortably: if you have to read it twice, so will your audience. Automatic tools make reasonable guesses here, but they split by time or character count, not by meaning, so a quick pass through the breaks is worth it.
Subtitles, captions and non-speech sound
In everyday use, "subtitles" and "captions" get mixed up. In the strict sense, subtitles carry the dialogue (often translated) for viewers who can hear the audio, while captions also describe what a viewer who cannot hear would miss: music, laughter, a door slamming, who is speaking.
That difference matters because automatic transcription only covers speech. The W3C Web Accessibility Initiative states that high-quality captions need skill in formatting, timing and including relevant non-speech audio, and that simple transcription is not enough. If your video will be watched by deaf or hard of hearing viewers, add sound cues by hand after the automatic pass, for example [applause] or [tense music].
This also pays off for people with no hearing loss at all. Many viewers watch with the sound off in public places, and captions keep them with you. If you cut sports or performance clips for scouting or promotion, the same logic applies: a short, captioned reel travels further. This guide to making football highlight videos shows how a well-structured clip gets seen.
Burned-in captions or a subtitle file?
You have two ways to deliver subtitles, and the right one depends on where the video goes.
| Burned-in (open) captions | Sidecar file (SRT or WebVTT) | |
|---|---|---|
| How it works | Text is rendered into the picture | Text lives in a separate file the player loads |
| Can viewers turn them off? | No | Yes |
| Styling control | Full: font, color, animation | Limited by the player |
| Multiple languages | One version per language | One file per language, same video |
| Best for | Short vertical clips, social posts, silent autoplay | Long-form video, websites, YouTube, courses |
| Editing later | Re-export the video | Edit the text file |
Which file format to choose
SRT is the plain, near-universal option: numbered cues, each with a start and end timecode and the text. It is accepted almost everywhere and opens in any text editor. WebVTT is its web-oriented relative, with support for basic styling, positioning and metadata, and it is the native format for HTML5 video players. When in doubt, export SRT, and add WebVTT if your own website player asks for it. Always check the upload page of the platform you use, because the accepted formats are listed there and sometimes change.
A small habit saves trouble: keep the SRT in the same folder as the final video, named with the language code (for example, launch-video.en.srt). When you add translations later, they stay organized.
Subtitles for Reels, Shorts and TikTok
Short vertical video is where automatic subtitles earn their keep, because most of it is watched without sound at first glance. A few rules help:
- Burn them in. Platform caption features differ and viewers cannot always toggle them, so styled, burned-in text keeps your look consistent.
- Stay out of the interface zones. Keep text inside the central area, away from the bottom strip where buttons and descriptions sit. Our vertical video editing workflow covers the 9:16 frame and safe areas in detail.
- Show a few words at a time. One to four words per cue, with the active word highlighted, reads fast and matches the rhythm of short clips.
- Keep contrast high. White text with a dark outline or a solid box holds up over any background.
- Still proofread. A wrong word in a big, animated caption is more visible than in a quiet subtitle line.
Translating subtitles with AI
Machine translation of a transcript is quick, but subtitles are not ordinary text. The Digital Entertainment Group's content globalization white paper notes that translation for subtitling has to respect reading-speed requirements, while translation for dubbing has to be adapted to support visual lip sync. Those are two different jobs with two different constraints.
In practice, that gives you a checklist for translated subtitles:
- Translate from a corrected transcript. Errors in the source language multiply in translation.
- Expect text expansion. Many languages need more characters than English for the same idea, so lines run long. Shorten rather than shrink the font.
- Watch idioms and humor. Literal translations of slang are the most common failure.
- Keep names consistent. Decide once how product and character names appear and apply it everywhere.
- Have a fluent speaker skim it. If the video is client-facing, legal or medical, a human review is not optional.
AI dubbing, voiceover and text to speech
Subtitles are not always the right answer. Some viewers prefer to listen in their own language, and some projects have no recorded voice at all. Three related tools cover those cases.
Text to speech and AI voiceover
Text to speech turns a written script into a synthetic voice. It suits explainers, tutorials, product demos and any video where you have a script but no time in a booth. The workflow runs in reverse of subtitling: write the script, generate the voice, then cut picture to the narration, and finally add captions from the same script, which means the text is already correct. If you work from a script, our guide to text to video animation shows how to turn it into moving scenes, and the explainer video animation guide covers narration-led formats.
Listen for pacing before you commit. Synthetic voices tend to flatten the end of sentences and stumble on acronyms, so spell tricky words phonetically in the script and add pauses with punctuation.
AI dubbing
AI dubbing replaces the original speech with a translated voice. The pipeline usually runs in four steps: transcribe, translate, synthesize a voice in the target language, and mix it back under or over the original track. The DEG paper's point applies directly: a dubbing script is not a subtitle script. It must fit the timing of the original speech, and when faces are visible it should suit the mouth shapes on screen. A translated line that runs two seconds longer than the original will either be rushed or spill over the next shot.
For YouTube and other long-form uploads, many creators publish the original video with translated subtitles first and add dubbed audio tracks only for their strongest markets. Subtitles cost less to produce and are easier to correct.
Lip sync dubbing and talking heads
When the speaker's face is on screen, mismatched mouth movement is what makes a dub feel off. Lip sync tools adjust the mouth to match the new audio. We explain the mechanics, and how to run it with AI, in our lip sync animation guide. For a talking head, a presenter or character who speaks straight to camera, the same approach lets you drive a face from a voice track instead of filming it.
What to review before you export
Run through this list on every video. It takes five minutes and catches nearly everything.
- Spelling of names, brands and titles, checked against the source, not memory.
- Numbers, dates, prices and units.
- Speaker changes, marked with a dash or a label when it is not obvious who is talking.
- Non-speech sound that carries meaning: music, alarms, laughter, off-screen voices.
- Sync at the start, middle and end. Drift often shows up late in a long file.
- Line length and breaks. No orphan single words on a second line.
- Profanity and sensitive terms, which engines sometimes mask, change or invent.
- Safe area and contrast on the actual device size, not just your monitor.
- The exported file, opened on the target platform, with sound off.
Our perspective: captions are part of the edit, not an afterthought
We build tools for people who make music videos, short films and ads, so we see the same pattern again and again. Creators treat subtitles as a final chore, and then the text collides with a lower third, the lyrics go out of sync after a late trim, or a name is spelled wrong in a title card. The fix is to plan for text from the first storyboard: leave room in the frame, decide whether the words are lyrics, dialogue or both, and lock the cut before you time anything.
Music is the hardest case. Sung words are the least reliably recognized speech, so for music videos we think it is better to start from the lyrics you already have than to ask an engine to guess them. Our lyric video tutorial follows that order: lyrics first, then vocals and timing. The same idea applies to any script-based video. Text you wrote is always more accurate than text a machine inferred.
We also think the way you pay for tools should match the way you work. A creator who needs subtitles on three videos this month and none next month should not carry a subscription for the quiet months. That is why we use prepaid credits rather than a recurring plan, and why we recommend you compare cost per finished video, not cost per month, when you choose any tool.
— simeon
Where Brix36 fits
If your video is a music video or a short film, the picture, the performance and the words all need to line up. Brix36 gives you section-based storyboarding, one-tap looks and an AI lip sync music video workflow that follows your lyrics, so the vocal performance and the visuals are built together instead of patched together afterwards.
You pay with prepaid credits rather than a subscription, so you only spend on the videos you actually make. See how credits work on the pricing page, or start from your song and build the video around it. If you are planning a narrative piece, the short film maker covers the same storyboard-first approach. To keep every project consistent with your own look and sound, Artist DNA stores that identity so you do not rebuild it each time.
If you are connecting several tools for transcription, voice and editing, our AI video workflow guide shows how to link them without redoing work at each handoff. And if you are showing your work to clients, a video portfolio with captioned samples makes it easier to view in any setting.
FAQ
How accurate are automatic subtitles?
Accuracy depends mostly on your audio. Clear, single-speaker speech with little background noise gets close to a usable draft, while music, accents, crosstalk and technical terms cause errors. Captions can misread words, context or sentence structure and drift out of sync, so review every file before you publish.
Can I add subtitles to a video for free?
Many online tools offer a free tier or trial credits, usually with limits on length, resolution, watermarks or export formats. Free editors and platform-native caption features also exist. Check what the free version exports (burned-in video, SRT or both) before you build a workflow around it.
What is the difference between subtitles and closed captions?
Subtitles present dialogue, often translated, for viewers who can hear the audio. Captions also describe non-speech sound such as music and sound effects, which is what the W3C recommends for accessible video. Closed captions can be switched on or off, while open captions are burned into the picture.
Is AI dubbing the same as translated subtitles?
No. Subtitles keep the original voice and display translated text, which has to respect reading speed. Dubbing replaces the voice, so the script must be adapted to fit timing and, when faces are visible, lip movement. Subtitles are cheaper and easier to correct, while dubbing suits audiences who prefer to listen.
Which subtitle file should I upload, SRT or VTT?
SRT is the safest default because nearly every platform and editor accepts it. WebVTT is the better choice for HTML5 players on your own website, since it supports styling and positioning. Check the upload instructions of the platform you are using for the formats it accepts.
Sources
- Auto-generated captions guidelines, Texas A&M University-Corpus Christi
- Captioning with speech to text, Microsoft Learn
- Content Globalization White Paper, Digital Entertainment Group (2024)
- Captions/Subtitles, W3C Web Accessibility Initiative