Podcast Captions and Subtitles That Actually Match What Was Said

The right way is a machine first pass followed by a human correction pass against the actual audio, then styling that fits where the caption will be watched. That order matters. Machines are fast and cheap at getting 90% of the words down. They are terrible at proper nouns, crosstalk, accents, and anything technical. Humans are slow at typing 9,000 words and very good at catching the four places the machine invented something.

So the workflow I run is simple. The episode gets transcribed automatically. Then the editor who is already cutting the episode reads the transcript while the audio plays, and fixes names, brands, numbers, and any line where two people talked over each other. Because that editor watched the whole episode anyway, this adds maybe 25 to 40 minutes on a 60 minute show, not a whole second job.

Nothing in my process is auto-cut by silence detection, and nothing is auto-captioned and shipped without a person reading it. That is the entire difference between a caption track and a caption track you would be happy to have quoted back to you.

After the words are right, styling comes last. Long form YouTube gets a clean SRT sidecar. Vertical clips get burned in captions sized for a phone. Those are different jobs and I treat them that way.

Auto captions fail on names and numbers because speech models predict the most likely word, and your guest's surname is almost never the most likely word. The model has heard "Anderson" a million times and your guest's actual name maybe never. Same with your product, your company, your city, and any figure said quickly.

The pattern repeats in a way that is almost funny once you know it. Numbers get flattened, so "one hundred and eighty thousand" becomes "180" or "one 80". Currencies drop, so "$40 million" turns into "40 million" with no unit. Hyphenated brand names get split. Acronyms get spelled as words. And crosstalk, where the host laughs over the guest's punchline, tends to produce one merged sentence that reads like nonsense.

The cost of this is not aesthetic. Captions feed search. YouTube reads them, podcast platforms read them, and AI answer engines lift sentences straight out of them. If your company name is misspelled in the caption track for 40 episodes, that is 40 episodes that do not surface when someone searches your company name.

This is also why I flag terminology up front. Before the first episode I ask for a short glossary: guest names, product names, recurring jargon, the way you spell your own brand. Ten minutes of your time saves an hour of mine every single week.

Use a separate SRT file for long form and burned in captions for vertical clips. That is the short answer and it holds up almost everywhere.

Long form episodes on YouTube should carry an uploaded SRT. Viewers who want captions turn them on, viewers who do not are not staring at a text bar for 70 minutes, and the platform gets a clean machine readable transcript for search. Burning captions into a 70 minute episode also means every future correction requires a full re export, which is a genuinely annoying way to live.

Vertical clips are the opposite. Most of that viewing happens with sound off in a queue, in public, at speed. If the words are not on screen the clip does not work at all. So clips get burned in captions, two to four words at a time, sized so they survive being watched on a cracked phone screen in daylight. I keep them in the middle third, above the platform's own interface junk at the bottom and clear of the face.

A third option exists and is worth knowing about: shipping both. Burned in for the clip, plus the corrected SRT for the parent episode, both coming from the same corrected transcript. That is what I do by default, and it is why the clips never disagree with the episode they came from. Consistency is free when the source is the same.

Corrected captions for the main episode are included in my editing rate, not billed as an extra line. Starter is $199 per episode for a single camera show at 1080p with one revision round. Studio is $299 per episode for 2 to 4 cameras at 4K with unlimited revisions. Both are priced per episode up to 60 minutes as the base, with up to 30 minutes cheaper and up to 90 minutes more.

The add-ons are where captioned deliverables get counted separately. Vertical clips are $25 each and come captioned. Show notes and chapters are $35. Rush 24 hour delivery is $75. Standard turnaround is a first cut back in 48 hours, and revisions come back the next working day.

Buying in blocks is cheaper. A 4-episode pack saves 15% and going monthly saves 25%. If you want the full breakdown it is on the pricing section rather than buried in a proposal.

Compare that to the usual shape of the market. A typical agency runs $300 to $400 per episode and then charges roughly $400 a month again for clips, which is where captions usually hide as a surcharge. Standalone caption vendors price per audio minute, which sounds cheap until you multiply it by a weekly 70 minute show. If you are weighing the automated route honestly, I have written that out at Opus Clip versus human clip editing.

A captioned sponsor read is a searchable sponsor read, and that is a number you can put in a media kit. When the brand name is spelled correctly in the caption track, that segment can be found, quoted, and cited. When it is spelled wrong, the read effectively did not happen as far as search is concerned.

This matters more than most hosts expect. Sponsors increasingly ask for proof of placement, and a timestamped chapter plus an accurate caption line is proof. It also makes the clip of that read reusable: the brand can post it themselves, captions already on, no extra work from you. Hosts tell us that handing a sponsor a ready captioned 40 second cut of their read is the single easiest renewal conversation they have.

So I treat sponsor segments as high priority lines in the correction pass. Brand spelling, offer code, and URL get checked character by character, because a discount code with one wrong letter is a dead link with a receipt attached. If you are structuring reads across a season, the mechanics live in sponsor integration.

One more thing worth saying plainly. Captions do not make a bad read good. They make a good read findable. The writing and the delivery are still yours.

The most common mistake is captions that are technically accurate and completely unreadable. Six lines of text at once, dumped on screen, in a font that fights the footage. The words are all correct and nobody can follow them.

After that, in rough order of how often I see it: captions timed to the transcript rather than to the speech, so they land a beat late all episode. Captions that sit under the platform interface on vertical clips, so the last line is permanently hidden. All caps everywhere, which reads as shouting and is measurably slower to scan. A different caption style on every clip, because a different freelancer made each one. And filler words left in verbatim, so the text is a wall of "um" and "you know" that the ear ignored but the eye cannot.

There is also the silent one: no line breaks at natural phrase boundaries. Splitting mid phrase makes a viewer reread. Break where the speaker breathed.

My fix list is boring and it works. Two lines maximum for long form, two to four words for vertical, break on phrase boundaries, one locked style per show, filler cleaned but meaning untouched. If you are setting a show up from scratch, the wider version of this is in the podcast launch checklist, and packaging the thumbnail to match the caption style is what my sister studio ClickTheory handles at 8 thumbnails for $250.

You do. Footage, project files, exports, transcripts, and caption files are yours, always, with no exit fee and no hostage situation. I keep a copy so I can work, you keep the originals, and if you leave you take everything in a form another editor can open.

I say this because the caption track is quietly one of the most valuable assets a show produces. It is the raw material for your show notes, your newsletter, your blog posts, your search visibility, and increasingly for whatever AI system someone uses to ask a question your episode already answered. A vendor who hands you a burned in MP4 and nothing else has taken that away from you without mentioning it.

So the delivery for every episode includes the corrected SRT alongside the video. Plain text transcript on request. No proprietary format, no login required to retrieve it, no per download charge.

The same principle applies if you are an agency reselling my work under your own name. The files belong to you and your client, my name is nowhere on them, and the caption style follows your brand rather than mine. That arrangement is spelled out under white label podcast editing. It works the same for a single host as it does for a studio running eight shows.