How to transcription and captions (step by step)

Open with a question the reader is already asking, then answer it bluntly.

I get emails every week from people who’ve just spent three hours typing out a one-hour meeting recording. They’re frustrated, their necks hurt, and the transcript still has “um” and “uh” everywhere. The blunt answer: you don’t need to type a word. You need the right engine doing the heavy lifting, then a quick pass to clean up the edges. I’m going to show you exactly how we stopped scheduling transcription meetings and started letting AI do the grunt work, without losing the details that actually matter.

Choosing the right transcription engine for your content

Not all engines are created equal. I’ve seen teams plug Whisper into a messy conference call and wonder why the speaker labels are missing halfway through. The first decision point is audio quality: clear single-speaker voice memos hit 95%+ accuracy out of the box, but throw in three people talking over each other, a cheap laptop mic, and a HVAC hum, and you’re looking at a much rockier ride.

If your primary use case is sales calls or internal team meetings, look for a tool that already knows the context. Some engines let you upload a custom vocabulary—product names, industry jargon, even specific employee names—and they’ll weight those words heavily during recognition. We started doing that last quarter and the “false positive” rate on our product names dropped from 12% to under 2% in a single afternoon.

Also think about language support. If you have remote teammates who regularly speak English as a second language, or you’re transcribing international client calls, make sure the engine handles non-native accents well. Most free demos will let you upload a short clip and see how it performs; I always take five minutes to test with my most common accent patterns before committing to a paid tier.

The step-by-step workflow that actually sticks

I’m going to lay this out as a numbered process because the first time I tried “just wing it,” I ended up with a disorganized text file and no idea where to start editing. Here’s the exact order we use now, and it takes about 10 minutes from upload to finished transcript.

1. Upload the file. Most platforms accept MP3, WAV, or M4A. If you’re pulling from a video platform like Zoom or Loom, check if there’s a one-click “export transcript” button—we save at least two clicks per meeting that way.
2. Run speaker diarization. This is the magic step that labels “Speaker A” and “Speaker B” instead of giving you a wall of text. Turn it on if you have more than one person on the call. If the tool can’t auto-detect, assign one person to listen and tag names as they go—it’s way faster than typing everything yourself.
3. Review for the 5% that matters. AI will mess up proper nouns, technical terms, and anything spoken under background noise. I skim the whole thing once, flagging anything that looks off, then play the audio for those specific 30-second chunks. Usually it’s just three or four quick fixes per hour of recording.
4. Add timestamps if you’ll reference them. If anyone on the team might want to jump to a specific moment—“what did they say about the pricing model at 12:45?”—keep the timestamps enabled. They turn a flat document into something searchable.
5. Export and store. We export as TXT or SRT depending on whether we need the file for captions or just a written record. Everything lands in a shared folder with the meeting date and topic in the filename, so searching later is instant.

The whole thing takes me about 8 minutes for a 30-minute meeting. If you’re still spending an hour typing, you’re using the wrong step or the wrong engine.

Why captions belong in the video player, not as a separate file

A lot of teams treat captions as an afterthought—a separate SRT file they upload to YouTube and call it done. I used to do the same, until I noticed that our internal demo videos had significantly lower completion rates when the caption file was hosted externally. The viewer had to click, wait for the file to load, and sometimes the sync was off by a few seconds.

Embedding captions directly in the video player changes the behavior. When the text is part of the player—especially on mobile, where auto-play without sound is the norm—people watch longer. We saw a 17% lift in average watch time last year after we switched to burned-in captions for our onboarding series. The trade-off is you can’t toggle them off, but for our use case (training videos watched on company laptops), that wasn’t a loss; it was a feature.

If you need both— toggles for accessibility compliance and burned-in for social—look for a platform that lets you layer both. Upload an SRT for the accessibility crowd, and enable the player’s built-in caption track for the “always-on” crowd. It’s a small config change that keeps everyone happy without doubling your workflow.

Common mistakes that kill transcription quality (and how to avoid them)

I’ve made all of these at one point or another, so I’m not guessing. These are the four things that consistently drag down accuracy, and the fix each time.

Mistake 1: Skipping the audio check. I’ve seen people upload a 20-minute recording where the speaker is two feet from a running fan, then complain the transcript is “useless.” Test your microphone or recording setup first. If you’re on a call, ask everyone to mute when they’re not speaking. Five minutes of prevention saves an hour of editing.

Mistake 2: Expecting perfect proper-noun handling out of the box. AI doesn’t know your company’s product names, acronyms, or the nickname your team uses for a competitor. Spend two minutes adding a custom vocabulary list before you run the engine. It’s a one-time setup that pays off every single file after that.

Mistake 3: Editing in a vacuum. Don’t read the text silently and change words based on “what sounds right.” Play the audio section you’re editing. AI often swaps homophones (like “quiet” for “quit”) or mishears technical terms; hearing the actual audio catches those every time.

Mistake 4: Forgetting speaker labels. A transcript without speaker tags is just a diary entry. If you’re sharing it with a team, make sure diarization is on, or add names manually in the first pass. It makes the difference between a document you can actually act on and one you ignore.

Comparing the top transcription tools for small teams

I’ve tried a handful of platforms over the last two years, and the “best” depends heavily on your volume and budget. Here’s the honest comparison from our experience, no sponsored fluff.

Otter.ai is the one most teams start with. It has a clean web interface, real-time transcription during live calls, and decent speaker labels right out of the box. The free tier gives you 300 minutes a month, which is enough to test the waters. Where it struggles is with background noise; if you’re recording in a coworking space or a home office with thin walls, the accuracy drops noticeably. We still use it for quick internal syncs, but anything client-facing gets a second pass.

Descript earns points for its built-in audio editor. You edit the text, and the audio moves accordingly—delete a word in the transcript, and the silenced section disappears from the audio file. That alone saves us at least an hour a week when we’re cutting demo reels. The transcription accuracy is solid (mid-90s on clear audio), and the custom vocabulary feature is easy to use. The price point is higher than Otter, so we only pull it out for video-heavy weeks.

Whisper (OpenAI) is the heavy hitter if you’re technical. It’s free if you have API access, and the accuracy on clean audio is genuinely impressive. The catch is you’re on your own for editing, speaker labeling, and formatting. We use it for bulk archive transcription—thousands of hours of recorded workshops—and batch-process the files with a simple script. If you have dev resources, it’s the cheapest per-minute cost around. If you don’t, the setup time outweighs the savings.

Rev is the old-school choice: human transcriptionists. I still recommend it for high-stakes legal or medical notes where a single missed word could be costly. It’s slower (24-hour turnaround on standard orders) and more expensive (roughly $1.25 per audio minute), but the accuracy is near 99% if the audio is clean. We keep a Rev account on standby for client contracts and anything with regulatory requirements.

Bottom line: if you’re doing under 10 hours a week of meetings, Otter or Descript will cover 95% of use cases. If you’re archiving legacy recordings or have the bandwidth to script a Whisper pipeline, the cost per minute drops dramatically. Try the free tiers first; I always do.

Exporting captions that work everywhere

Once you have a clean transcript, the next question is format. Different platforms want different files, and picking the wrong one means you’re re-uploading or converting later. Here’s what we’ve learned about making captions portable.

SRT is the most universal. You’ll find it on YouTube, Vimeo, most social video players, and even some HTML5 video embeds. It’s just plain text with timecodes; no styling, no speaker names, just “when does this line appear and disappear.” If you need one format that works everywhere, SRT is it.

VTT (WebVTT) is the HTML5 standard. It supports things like speaker identification, background speaker labels, and even small formatting cues like bold or italics if you need them. Most modern video players accept VTT without complaint, and it’s a little more flexible than SRT if you want to add minor styling. We use VTT when we’re embedding videos on our own site and want the player to display speaker names in the caption track.

Burned-in (open captions) are the lines permanently part of the video frame. No file to upload, no player setting to toggle. Just a video with text already on it. Like I mentioned earlier, this is our go-to for social media and internal training where we want 100% compliance without relying on the viewer’s player settings.

Caption files with speaker labels are a niche but useful format if you’re producing a multi-person interview or a panel discussion. Some platforms let you upload an SRT or VTT where each line includes “Speaker A:” at the start. It makes the transcript readable even when sound is off, which is a nice accessibility win.

The rule of thumb: export an SRT for broad compatibility, a VTT if you’re building your own player, and consider burned-in for anything destined for Instagram, TikTok, or X. We keep all three versions in our asset folder now; it adds maybe 30 seconds to the export step, and saves us from playing “guess the format” later.

Conclusion: stop typing, start editing

If you took away one thing from this, let it be this: stop treating transcription like data entry. The days of listening to a recording and typing every word are gone, and honestly, they should stay gone. The right engine does the heavy lifting in minutes, and your remaining job is a quick skim to catch the three or four terms the AI got wrong. That’s not “editing” in the traditional sense; it’s quality control, and it takes a fraction of the time.

The real win is what you do with that hour you just reclaimed. We’ve used it for actual selling, for deeper client follow-ups, and for finishing that strategic doc that always gets pushed to the back of the queue. If your team is still scheduling “transcription meetings” on the calendar, stop. Set up the automation once, let it run in the background, and start using the output instead of building it.

We’ve been using Recordly to handle the heavy lifting for our meeting notes and video captions, and it’s been the easiest way to keep everything accurate without dedicating a full hour per recording. It handles speaker diarization, custom vocabulary, and exports in the formats you need without the usual faff. If you’re tired of the typing treadmill, give it a spin—you might be surprised how little manual work actually remains.

Keep reading

Playbook

Auto-zoom in screen recordings: what it is and why your demos need it

A hundred random links can do less than five relevant ones. Here's what makes a backlink pass real ranking value in 2026 — and what makes it worthless or worse.

Sofia Meyer October 13, 20262 min read

Playbook

How to choose a screen recorder for product demos in 2026

A hundred random links can do less than five relevant ones. Here's what makes a backlink pass real ranking value in 2026 — and what makes it worthless or worse.

Sofia Meyer October 16, 20262 min read

Playbook

Video chapters: the timestamp detail that keeps viewers watching

A hundred random links can do less than five relevant ones. Here's what makes a backlink pass real ranking value in 2026 — and what makes it worthless or worse.

Sofia Meyer October 13, 20262 min read

Explainer

Dofollow vs nofollow in 2026: what still passes value

The link attributes changed, Google's treatment changed, and a lot of advice is out of date. Here's what dofollow, nofollow, sponsored and ugc actually do now.

The Recordly Team October 10, 20262 min read