DEV Community

Andriy Kuzmenko
Andriy Kuzmenko

Posted on Originally published at andriyvk.github.io

YouTube subtitles in Ukrainian - transcript and translated SRT by API

#youtube #python #ai #api

Plenty of tools fetch a YouTube transcript. Fewer give you subtitles in Ukrainian, timed to the video, ready to
load into a player or an editor. This guide shows one call that returns the original transcript, its Ukrainian
translation and both as SRT/VTT files. The output below is from a real run on 7 November 2026.

Disclosure: I built the tool used here, Video Transcript & Translation
on Apify. The source code is open: AndriyVK/-apify-document-tools.

The call 

from apify_client import ApifyClient

client = ApifyClient("<YOUR_APIFY_TOKEN>")
run = client.actor("ai_admin/social-video-transcript").call(run_input={
    "videoUrls": ["https://www.youtube.com/watch?v=jNQXAC9IVRw"],
    "translateTo": "uk",      # Ukrainian; 20 more languages are available
    "subtitles": True,        # SRT and VTT, original and translation (free)
})
for item in client.dataset(run["defaultDatasetId"]).iterate_items():
    print(item["title"], item["transcriptSource"])
    print(item["translation"]["text"])
    print(item["translation"]["srtUrl"])
Enter fullscreen mode Exit fullscreen mode

The same list takes TikTok, Instagram Reels, Facebook, X and Vimeo links.

The real output

The video is the first one ever uploaded to YouTube, «Me at the zoo» (19 seconds).

title: Me at the zoo   language: en   transcriptSource: uploaded-captions   seconds: 18.2
Enter fullscreen mode Exit fullscreen mode

The Ukrainian SRT file (first three cues):

1
00:00:01,200 --> 00:00:03,360
Гаразд, тож ми тут, перед слонами

2
00:00:05,320 --> 00:00:07,970
круте в цих тварин те, що у них справді...

3
00:00:07,970 --> 00:00:12,620
справді дуже довгі хоботи
Enter fullscreen mode Exit fullscreen mode

Each translated cue keeps the timing of the original one, so the file lines up with the video without any editing.
The translation is natural Ukrainian («довгі хоботи», not a word-for-word copy), and even the elephant sound
(baaaaaaaaaaahhh!!) became (бааааааааа!!).

Captions or speech recognition

  • The video has captions (uploaded or YouTube's automatic ones): they are used. Fast, any length, one flat price. That was the case above: uploaded-captions.
  • No captions: the audio is recognised with Whisper large-v3, which handles Ukrainian well. Then you pay per minute, and maxDurationMinutes (60 by default) keeps you from transcribing a four-hour stream by accident.

TikTok without captions: four real runs

TikTok videos usually have no caption track, so here the audio itself is recognised. Four public videos, one run on
7 November 2026, with Ukrainian translation switched on ("translateTo": "uk"):

run_input = {
    "videoUrls": [
        "https://www.tiktok.com/@nasa/video/7698805564884733198",
        "https://www.tiktok.com/@bbcnews/video/7699782697031404822",
        "https://www.tiktok.com/@hromadske.ua/video/7704723523511635221",
        "https://www.tiktok.com/@hromadske.ua/video/7704712902082874645",
    ],
    "translateTo": "uk",
    "subtitles": True,
    "includeSegments": True,
}
Enter fullscreen mode Exit fullscreen mode

All four came back with transcriptSource: speech-recognition, none failed. The excerpts below are copied from the
output as they are, mistakes included.

1. NASA, English, 45 seconds, 11 segments. Clear studio speech, recognised without a single error I could spot:

00:14  Viper very much acts as a scout for Moonbase.
00:19  We've already built three mock-ups, actually two of them within three weeks.
00:37  Our goal is to enable humans to be operating on the surface of the moon on a permanent basis.
Enter fullscreen mode Exit fullscreen mode

Ukrainian translation of the last line: «Наша мета - забезпечити можливість для людей працювати на поверхні Місяця на
постійній основі.»

2. BBC News, English, 35 seconds, 5 segments. A weather presenter outdoors, in a storm:

00:00  The full force of Storm Austin has arrived here in Newquay and Cornwall.
       We are now seeing gusts of wind of 17, even 18 miles an hour.
00:16  These strong winds have the potential to bring down branches and trees. There could well be power cuts.
00:21  Now the strong winds will move their way a bit further east with drought in the morning ...
Enter fullscreen mode Exit fullscreen mode

This is the hard case, and it shows. The storm is called Austen (the video's own hashtag), not Austin. Gusts of
17-18 mph would not bring down trees, and the Met Office forecast for Cornwall that day was 60-70 mph, up to 80 in
exposed places. I cross-checked by running the same video again with "quality": "fast" (Whisper large-v3-turbo):
it heard «70, even 80 miles an hour» and «throughout the morning», so «drought», which the translation faithfully
turned into «посуха», is a mishearing too. The second model made its own mistake in the same sentence («just a room
of» instead of «gusts of wind of»). Wind noise is what breaks numbers and short words, so check figures in such videos.
Since this test the tool does it for you: with "crossCheck": true it recognises the audio with both models and lists
the places where they disagree, with times, figures flagged ("numbers": true). On this video the list starts with
«gusts of wind of 17 even 18» against «just a room of 70 even 80», then «with drought in» against «throughout». The
cross-check doubles the recognition price: $0.02 instead of $0.01 for a video up to three minutes.

The same option is in Audio & Video Transcriber, for your own files:
meetings, interviews, calls. There it caught something worse than a misheard number. On a 45-second test recording,
where one paragraph is read three times, the main model wrote a single word, «The», for the stretch from 0:18 to 0:25
and lost two sentences of the second reading. The cross-check listed that stretch with the full text the second model
heard there. Without it the gap is easy to miss: the transcript still reads smoothly. Every minute is charged twice
with the cross-check, $0.01 a minute instead of $0.005.

3. hromadske, Ukrainian, 74 seconds, 14 segments. A performer in an interview, conversational speech:

00:00  Я себе ловлю на тому, що мене чіпляє все, що з Росією пов'язано.
00:10  У 22-му році я поїхав в Естонію.
00:15  І я в аеропорту підхожу до гейту і чую позаду мене чокання, акання.
00:51  І буде відчуття несправедливості, і буде нас це тригерить постійно ...
Enter fullscreen mode Exit fullscreen mode

Even slang like «тригерить» and «чокання, акання» came through. The language was detected as Ukrainian, so the
translation was skipped with the warning the video is already in the target language and no translated minutes were
charged.

4. hromadske, Ukrainian, 91 seconds, 7 segments. A calmer interview about the Ukrainian language:

00:00  Захист мови є частиною національної безпеки.
00:08  З того часу, коли росіяни зросифіковували українців, вони діяли страхом, примусом ...
00:19  А з іншого боку, у нас демократична країна, демократичні методи українізації суспільства.
Enter fullscreen mode Exit fullscreen mode

One word is wrong: «зросифіковували» should be «русифікували». The rest reads like a clean transcript. The long
segments (up to 20 seconds) follow the speaker's sentences; for subtitles on a phone screen you may want to split them.

What the run cost a user on a paid plan: four videos at $0.01 plus two started minutes of translation at $0.007, so
$0.054 in total ($0.017 for each English video with translation, $0.01 for each Ukrainian one). Videos up to three
minutes cost the same flat $0.01 with speech recognition as with captions.

Price

On a paid Apify plan: $10 per 1,000 videos with captions, plus $7 per 1,000 minutes translated. The video
above: one video plus one started minute of translation, $0.017. On the free plan: $0.025. A YouTube video without
captions is $10 per 1,000 minutes instead. Failed videos are free.

Limits

  • Private videos, videos that need a login and live streams are not supported.
  • YouTube sometimes refuses downloads from servers; such a video gets an error field and is not charged.
  • Speakers are not labelled.
  • The translation is done by a language model: good for watching and searching, but read it before you publish it as official subtitles.
  • Transcribe and translate only videos you are allowed to use.

From an AI agent

https://mcp.apify.com?tools=ai_admin/social-video-transcript gives Claude, Cursor or any MCP client one tool that
reads a video link and answers in text.

Originally published at andriyvk.github.io.

Top comments (0)