How to Get a Transcript From a YouTube Video (Plus Reels and TikTok) — With Timestamps, SRT, and VTT

The end result isn't one thing. You get readable text plus downloadable SRT and VTT with word-level timings — and three different shapes of that text, each for a different job. Plain paragraphs are for quotes, a post, a description, a script draft: places where time doesn't matter. Timed captions are for anything headed into a player, an editor, or a platform — every line knows when it starts and when it ends.

The third shape is chapter timestamps. Same text, but marked with points like 00:00, 4:12, 11:38 that drop into a long-video description so viewers can jump between sections. Handy when the transcript already exists and the video is long enough to carve into chapters for the description or for clipping.

• Plain text (TXT) — quotes, posts, descriptions, a script draft. No timing attached.
• Timed captions (SRT/VTT) — anything that goes into a player, a cut, or a platform.
• Chapter timestamps — the long-video description, so viewers skip to the part they want.

From here it's all about the platform. Some hand the text over ready-made and you just grab it; others force you to recognize it from the audio. That split is the next section.

YouTube generates auto-captions on its side and stores the result, which is why a video has a transcript panel and a downloadable caption track. You can grab that track — if the creator didn't disable it and if YouTube actually finished generating it. Sometimes there's nothing to grab: the upload is fresh, the spoken language isn't supported for recognition, or the creator switched captions off.

Reels and TikTok have no such track. They paint captions over the video, but that's part of the player, not a file you can pull. So for those platforms — and for any YouTube video with captions off — you grab the audio and recognize it yourself.

The practical difference matters. Platform auto-captions are someone else's tool: you don't control the language, the line breaks, or the accuracy. Your own recognition is a result you can check, fix, and export in the format you need. Here's how it shakes out platform by platform.

Simple order of operations: check for existing captions first, recognize the audio only if they're missing or read like garbled soup. That way you don't redo work the platform already did for you.

1. 1

   Copy the video URL

   Grab the address from your browser or the app. youtu.be/X and watch?v=X&t=90 are the same video — recognition knows that and won't do the work twice.

2. 2

   Paste it into the transcript field

   Open the tool and drop the link in. A channel or profile URL gets rejected — there's no such thing as a channel transcript, only a specific video. A media-library file works too, as long as the video or audio is under an hour.

3. 3

   See what's there

   If the video has an existing caption track, the text shows up fast and near-clean. No track, and recognition kicks in — the in-house model first, a cloud model only if the in-house one couldn't handle it.

4. 4

   Grab it in the format you need

   On screen: paragraphs, with SRT and VTT to download and hand-off buttons to send the text into a script or the link into clipping. Paragraphs are split by a blank line, so the text reads right away instead of sitting as one wall.

5. 5

   Check a seam at a pause

   Scroll to a spot where the speaker took a breath or changed topic. A line should close on that pause, not in the middle of a word. If everything's breaking wrong, re-recognize it.

One thing worth knowing about the line splitting. Cues break by words, not by recognition chunks: a line closes at the end of a sentence, at a pause of 0.8 seconds or more, at 84 characters (two 42-character lines — the ceiling in Netflix's caption guide), and at 6 seconds max. Without that, a half-hour conversation would come out as half-line cues that flicker in the player and don't read.

No ready track here, so there's one path: pull the audio and recognize it. A link works for vertical clips and standard video alike — the type gets detected automatically. If the clip was cut from a longer video, recognition still only gives you the text of that clip, and timings count from the start of the clip, not the start of the source.

Reels has its own quirk. The direct link sometimes won't return, and the pull falls back through the Instagram source. That doesn't always work — a private account or changed page markup can break it. If a clip won't pull, download it through the official export or the app and recognize it from the file. Same result; you're just bringing the source by hand.

1. 1

   Copy the Reel or TikTok link

   A plain public link works — no shortened URLs, no redirects. If the link points at a profile instead of a clip, the field rejects it.

2. 2

   Paste it and run recognition

   The audio pull has its own method per platform, then it's the same recognition chain as YouTube. Text shows up in paragraphs; timings count from the start of the clip.

3. 3

   If the link won't return, bring a file

   Download the clip and upload it to the media library. Video or audio up to an hour is accepted — the audio gets pulled to mono 16 kHz, which is plenty for speech.

4. 4

   Grab the text and decide what's next

   From the text you can build captions for the same clip, a quote for the caption, or a cut into clips — a vertical video is already set up to cut by speech.

This is exactly what the Transcript tool is for: one field takes a link from YouTube, Shorts, Reels, TikTok, or VK and returns the text in paragraphs plus SRT and VTT. Those same word-level timings then feed hooks, captions, and clips.

All three carry the same text; the difference is what's attached to it. TXT is paragraphs separated by a blank line — clean material for a quote, a post, an article, a script draft. There's no timing in it at all, which is exactly why it's useless for captions.

SRT is the caption standard: a numbered cue, a start and end time, the text line, a blank line, the next cue. Editors, players, caption tools, and most platforms read it. VTT is SRT's web cousin — same structure, but with a header and web settings, so it's easier to attach to a player on a site or in an app.

Cues inside SRT and VTT split by words, not by recognition segments — a segment can run half a minute, and no one reads that as a single cue. So a phrase closes at the end of a sentence, at a pause, by line length, and at six seconds max. Overlapping cues get trimmed too, otherwise the player shows two subtitles at once and the text jumps.

Recognition runs on the in-house model and costs nothing — in Kutto the transcript doesn't draw from your balance. If the in-house model can't handle a clip, a cloud step kicks in, and that one counts against your plan's AI volume as "AI Text." There's no per-item price line for the transcript — you don't pay per transcript. For current plan limits, see the pricing page.

There's no fixed seconds-per-minute figure — it depends on the queue and which model picks up the job. A short clip returns noticeably faster than an hour-long interview, and downloading the source by link is its own step that takes its own time.

Repeat runs are cheaper than the first. The result is cached by the video's key, so running the same link again returns from cache instead of recomputing anything. For a file, the file itself is the key — same asset, same text. A failed job doesn't land in the cache, so you can re-run it. And that result isn't just text: captions, hooks from the opening lines, description timestamps, and a cut into clips all get built from it.

Recognition is only as good as the audio under it. When music sits louder than the voice, or there's noise, echo, or a recording made through a speaker, the model starts inventing words where it can't make them out, and you'll be fixing text by hand. When two people talk at once, lines get mixed and, again, no speaker labels come out — you add those yourself.

Heavy accents, rare terms, names, and acronyms come out wrong on a regular basis. It's cheaper to check those up front than to catch them later in finished captions. And it flat-out won't work where there's no speech at all: a clip that's only music and on-screen text has nothing to transcribe.

No accuracy benchmark needed. Take the fastest-speech stretch — the first 30 seconds, where the speaker rushes, breathes, and jumps between topics — run it on its own, and read the text out loud next to the player. Watch the seams, not the typos: a line should close on a pause, not mid-word.

Then hit the three things that break most often. Proper nouns and names — check them against the video, not your memory. Numbers and dates — recognition regularly hears "fifteen" where the speaker said "fifty." Ragged bits at the very start and end, where someone talks over music — those are the first candidates for a fix.

• Verify names, brands, and terms against the video — recognition mangles those most.
• Scan the spots where two voices overlap: that's where lines get scrambled.
• Listen to numbers, dates, and amounts — those you hear, not read.
• Confirm cues don't overlap each other, or the subtitle will jump in the player.

Before you download anything, run these four points. They take about a minute and catch most of what otherwise forces a redo later.

• Format matches the job: TXT for quotes and posts, SRT or VTT for captions and editing, timestamps for the description.
• Names, terms, and numbers checked against the video, not memory.
• Cue seams are clean: lines close on pauses, no overlaps.
• You know where it goes next: into a script as raw material, or into clipping as clips and captions.

Text is only half the job. Most people turn it into something that ships: captions over the clip, clips from a long recording, or chapters for the description. Kutto takes a link or a file and returns the transcript with word-level timings — enough to cut by speech instead of eyeballing it.