TikTok Transcript Scraper: Video to Text, Captions and SRT

Turn TikTok videos into text, without an account or API key. Paste video links (short vm.tiktok.com links too) or give an account, and each video comes back with its transcript as plain text and as timestamped segments, its language, optional SRT and WebVTT subtitle files, and the video's metadata (account, date, views, likes, comments, shares, saves, caption, sound).

We use TikTok's own captions when the video has them (written by the creator or automatic): they are returned as they are, instantly, and cost no speech-to-text minutes. Otherwise the audio is transcribed by AI. Music-only videos come back as a free no_speech record instead of a made-up transcript. You pay only for transcripts delivered.

What we measured

On the Apify platform on 8 October 2026, with the default settings, on 42 inputs: links taken from Reddit posts that shared TikTok videos (most of them vm., vt. and /t/ short links), three accounts and a video ID:

• 26 transcripts in 5 languages (English, Japanese, Spanish, Portuguese, Chinese): 22 from TikTok's own captions (no minutes billed), 4 from speech-to-text.
• 15 videos without clear speech or without a video (music, loops, uncertain language, photo slideshows): free missing records.
• 5 failures with their reason, all free: deleted, private and age-restricted videos, an expired short link, an account that does not exist.
• 46 records in 93 seconds. The default input (one video) runs in about 7 seconds.

Running the same input again (13 records: damaged links, music videos, an account, unknown accounts) gave the same verdict on every record.

Speech-to-text runs on third-party AI providers: OpenAI's Whisper models (large-v3 and large-v3 turbo) run by Groq for most videos, and ElevenLabs Scribe, a stronger model, to settle the language when hints disagree and as a backup. The audio is sent to them only to be transcribed. ElevenLabs stores each transcript on its side: we delete it as soon as we have read it and check that it is gone (we read it again and get "not found"); if a deletion cannot be confirmed before the run ends, the run log says so.

On the videos that have both, TikTok's automatic captions and Whisper differ by about 7 % of words (mostly punctuation and names; measured on 23 videos). Whisper is known to invent short phrases on music ("Thank you.", "Music", "Bye."); we measured that on 170 videos and filter it: a phrase the model is not confident about, a known phantom phrase, or a few words spread over a long stretch of music are not speech. A video whose speech is all filtered out is missing with missingReason: "no_speech", free. We never sell part of a video as its transcript: when the filter has to leave out more than 20% of the speech (the model looped or was unsure over that part), the video is missing with missingReason: "incomplete_transcript", free. Loops are measured per character, so Hindi, Urdu, Arabic or Tamil text is judged like English (until 8 October 2026 a byte-based measure dropped correct Hindi sentences as "loops"). Whisper's output is also capped for each 30 seconds: dense speech in some scripts (Gujarati, Telugu, fast Hindi) loses the end of each 30 seconds without a trace. We detect it; such a video goes to a second, stronger model, and is never charged as a fragment.

TikTok's own captions pass the same checks before they are used: captions that are only a sound tag ("[Music]"), a lone phantom phrase ("Thank you."), a loop, or run longer than the video (captions of another video) are not sold; the audio is transcribed instead. A sound tag inside a caption line ("[Applause]", "♪") is removed from the text. Captions with no language are not used either.

What you can transcribe

Input	Examples	What you get
Videos	https://www.tiktok.com/@natgeo/video/7694024577701137677, 7694024577701137677	that video
Short links	https://vm.tiktok.com/…, https://vt.tiktok.com/…, https://www.tiktok.com/t/…	the video they lead to
Other forms	m.tiktok.com/v/….html, /embed/v2/…, /player/v1/…	that video
Accounts	@natgeo, https://www.tiktok.com/@natgeo	the account's latest videos

Mix them in one run. A video found in two inputs (a link and a short link, or a link and an account) is transcribed and charged once.

Accounts: we read the list TikTok publishes for embedding an account on other websites: about 10 of its latest videos (pinned ones first), without signing in. TikTok's full video list needs signed requests that we do not make; for older videos, paste their links. Photo slideshows do not take the place of a video, and when the list holds fewer videos than maxVideosPerAccount, the status says so.

Damaged links are refused, never guessed. A video link with a cut or extra character (…/video/76941., an empty /video/) gets a free invalid_url record; it is never read as the account in the link. An account is only @name or tiktok.com/@name with nothing after it.

Languages

TikTok's own automatic captions are used when they match the language of the post; when they do not (for example English speech captioned by TikTok's Japanese recognition), the audio is transcribed instead. For speech-to-text, we do not trust a single language detection when anything says otherwise. Hints: the caption's script (Devanagari, Hangul, kana…), Hindi written in Latin letters, and the language TikTok gives the post (English is not a hint: Whisper only translates into English, so it cannot catch a translation). When a hint and the detection disagree, a stronger speech-to-text model settles the language and transcribes the video (a Korean song that the fast model "translated" into English comes back in Korean); Hindi, Urdu and Punjabi go to it too. When the ElevenLabs referee is not available for a video, Whisper large-v3 settles the language and must agree, or nothing is sold. The referee is paid per use, so a run calls it only once the transcripts it has delivered cover the cost: in a larger run such a video waits for the others; sent alone, or in a run that delivers little, it can come back as a free uncertain_language record. If the language still cannot be settled, the video comes back as a free uncertain_language record, never as a charged guess and never as a translation. Hindi and Urdu are one spoken language in two scripts: the caption's script decides.

language is a preference, never an instruction: it changes neither the models used nor the number of passes, and never forces a translation (requestedLanguageFound: false when the video is in another language). With language set to Hindi, Urdu or Punjabi, a video that Whisper hears as English (with no caption or TikTok language saying otherwise) comes back as a free uncertain_language record: on those languages Whisper sometimes writes an English translation instead of what was said.

Output

One record per video. A transcript (status: "ok"), shortened here:

{
  "status": "ok",
  "platform": "tiktok",
  "videoId": "7694024577701137677",
  "url": "https://www.tiktok.com/@natgeo/video/7694024577701137677",
  "author": {
    "username": "natgeo",
    "fullName": "National Geographic",
    "id": "6780344874811442181",
    "isVerified": true,
    "url": "https://www.tiktok.com/@natgeo"
  },
  "caption": "The shoebill's prehistoric appearance makes it unique among living birds, …",
  "hashtags": null,
  "publishedAt": "2026-10-07T20:25:15Z",
  "viewCount": 6529,
  "likeCount": 529,
  "commentCount": 6,
  "shareCount": 5,
  "saveCount": 18,
  "sound": { "title": "original sound", "author": "National Geographic", "isOriginal": true },
  "durationSeconds": 105,
  "language": "en",
  "transcriptSource": "captions",
  "isAutoGenerated": true,
  "hasSpeech": true,
  "transcript": "Did you know shoebills attack their own siblings? Five surprising facts about the shoebill. …",
  "segments": [
    { "start": 0.06, "end": 2.94, "duration": 2.88, "text": "Did you know shoebills attack their own siblings?" },
    { "start": 2.941, "end": 5.061, "duration": 2.12, "text": "Five surprising facts about the shoebill." }
  ],
  "segmentCount": 37,
  "wordCount": 268,
  "characterCount": 1671,
  "billedMinutes": 0,
  "input": "@natgeo",
  "inputType": "account",
  "position": 3,
  "scrapedAt": "2026-10-08T00:10:12.607Z"
}

A video without a transcript comes back as a free record that holds the reason and nothing else (no account, counts or caption: those come with transcripts):

{
  "status": "missing",
  "platform": "tiktok",
  "videoId": "7693616413134703886",
  "url": "https://vm.tiktok.com/ZN8kbSguY/",
  "missingReason": "photo_post",
  "message": "this is a photo post (slideshow), not a video",
  "input": "https://vm.tiktok.com/ZN8kbSguY",
  "inputType": "video"
}

Fields that need a word of explanation

• transcriptSource: captions (TikTok's own captions; isAutoGenerated: false when the creator wrote them) or speech_to_text (AI transcription of the audio). Turn off useTikTokCaptions to send every video to speech-to-text.
• status: ok (transcript, charged), missing (the video exists but has nothing to transcribe: no_speech, incomplete_transcript, no_audio, photo_post, or too_long past your limit; free) or error (the link could not be read; free).
• errorCode and retryable: not_found (deleted video or expired short link), private, age_restricted, region_blocked, not_a_video (a short link that leads to an account), invalid_url, account_not_found, account_private, account_empty are final; unavailable, blocked, network, transcription_capacity and transcription_failed are worth retrying later (retryable: true).
• language: ISO 639-1 code of the language spoken and transcribed. We never translate: if you set language and a video is in another language, you get the language actually spoken, with requestedLanguageFound: false.
• filteredSegments: segments of the raw speech-to-text output left out of the transcript as non-speech (a "Thank you." over silence, a "[Music]" tag, a model loop); minutes are billed on the video's length as usual. A model loop or an unsure passage that held more than 20% of the speech makes the video a free incomplete_transcript record instead.
• billedMinutes: started minutes billed as speech-to-text for this transcript; 0 for TikTok's captions.
• segments: start, end and duration in seconds. includeSrt and includeVtt add ready-to-save .srt and .vtt files.
• input and position: which of your inputs produced the record, and the rank of the video in an account's list. Records are written as each video finishes: sort by input and position for the source's order.

The dataset has three views: Overview, Transcripts and Video details. For a spreadsheet, export the Transcripts view as Excel (XLSX): one row per video, and Excel treats every cell as text. The full record with segments does not fit a spreadsheet (one column per segment), and a CSV opened in Excel may read a caption that starts with =, +, - or @ as a formula: open CSV files as text, or prefer XLSX or JSON. Excel also cuts a cell after 32,767 characters (about 35 minutes of speech). Control characters and text-direction marks are removed from captions and transcripts. The run's summary is the OUTPUT record of the default key-value store: per input, videos, transcripts, missing, errors and duplicates, plus speech-to-text and network statistics.

Input examples

The latest videos of two accounts and a short link:

{
  "urls": ["@natgeo", "https://www.tiktok.com/@bbcnews", "https://vm.tiktok.com/…"],
  "maxVideosPerAccount": 10,
  "includeSrt": true
}

Every video through speech-to-text, in Spanish, text only:

{
  "urls": ["https://www.tiktok.com/@…/video/…"],
  "useTikTokCaptions": false,
  "language": "es",
  "transcriptFormat": "text"
}

Pricing

Pay per result: $0.001 per transcript, plus $0.001 per started minute only when the video is transcribed by speech-to-text, plus a start fee of $0.003 per GB of memory per run ($0.003 at the default 512 MB). A video with TikTok captions costs $0.001 whatever its length; a 90-second video without captions costs $0.003. Videos without speech, incomplete transcripts, photo posts, deleted or private videos, errors and duplicates are free. Platform usage is included.

Apify plan	Per transcript	Per started minute of speech-to-text
Free	$0.001	$0.001
Bronze	$0.0009	$0.001
Silver	$0.0008	$0.001
Gold and above	$0.00075	$0.001

Your maximum charge per run is never exceeded: the length of each video is read before it is transcribed, and the run stops at the last transcript it can charge and says so in its status. maxItems caps the transcripts of a run: ask for 10, get 10.

Free records are a by-product, not a product: a run stops, and says why, once 100 videos or inputs (an unknown account counts) plus 2 per transcript have given no transcript, or once 15 minutes of audio plus 2 per billed minute have been transcribed for nothing (music-only videos); audio we could not send because our capacity was used up does not count. maxDurationMinutes (default 30, at most 60) skips longer videos before anything is transcribed. A video with TikTok captions reserves only the transcript price, and a single video that would cost more than what is left of your maximum charge gets a free over_budget record while the run goes on with cheaper ones.

Use cases

• Trend and content research: what creators in a niche say, searchable and quotable.
• Repurposing: turn your own TikToks into captions, posts, newsletters and subtitle files.
• LLM pipelines: summaries, hooks, scripts and topic analysis across hundreds of videos.
• Brand monitoring: what is said about a product in the videos you collect elsewhere.

Limits

• Speech or music. The speech-to-text service does not return exactly the same text on every pass over the same video; on clear speech it does not matter, on music it can. So we keep a margin: a video is charged only when its speech is clear (enough words, good confidence, no loops such as "wrestles, wrestles, wrestles"); everything in between comes back as a free no_speech record. Clearly sung lyrics still count as speech. A song (another account's sound, or a track TikTok recognises as a released song even when it is posted as "original sound") is charged only when the lyrics are clear and extended: on two Whisper passes that agree, or on the single pass of the ElevenLabs model (it returns the same text for the same file, so a second pass would prove nothing).
• "Original sound" of another video (duets, stitches, reused sounds): a sound is used as the video's own audio only when its author is the video's author, it was created with the video and it has the same length; otherwise we take the video's own sound track.
• Accounts: about 10 latest videos (what TikTok's embed shows). No hashtags, sounds, search or comments.
• Private and age-restricted videos need a signed-in account; we do not use accounts, so they come back as free error records. Some videos are not shown in the region we read from (region_blocked).
• Challenge pages. TikTok sometimes answers with a bot check instead of the video (about 2 pages in 100 in our tests). We never solve those checks: the page is asked again automatically, and only then the video gets a free blocked record (retryable: true).
• Speed. Our speech-to-text capacity is shared fairly between runs and paced: count about 3 seconds per video without captions on large runs. If it is temporarily used up, videos that need it get free transcription_capacity records (retryable: true, "temporarily used up, retry later") while videos with captions keep coming; the status says when to retry. These records never count against your run's limits and never make it fail.
• Some short links (vm.tiktok.com/ZN9D…, 13 characters) lead to TikTok's home page even in a browser: they come back as not_found.

Memory and speed

512 MB is the default and the best setting: the run waits on speech-to-text, not on memory, and the Actor uses under 100 MB. More memory does not make it faster (the maximum is 1 GB).

FAQ

Is it legal to scrape TikTok transcripts? This Actor reads only public videos, the way a visitor who is not signed in sees them, solves no bot check and accepts no TikTok terms on your behalf. Transcripts reproduce what people say, which may be protected by copyright and may contain personal data: check that your use (research, analysis, quotation, accessibility) is allowed where you are, and handle personal data under the rules that apply to you (GDPR in Europe). We keep nothing after the run; the transcripts ElevenLabs stores are deleted during the run and the deletion is checked (the run log says so if one could not be confirmed).

Do I need a TikTok account or any setup? No. Keep the default connection: its cost is included in the price. If you set your own connection and it refuses everything, the run stops at once with a free proxy_failed record per input.

Why "captions" for one video and "speech_to_text" for another? About 4 TikTok videos in 10 have captions (63 of 150 in our sample); we return those as TikTok made them. The others go through speech-to-text. billedMinutes says what each video cost in minutes.

What happens when TikTok changes something? The Actor runs every day on known videos, so we see a change before you do. A run that fails on our side ends as Failed and charges no transcript.