Bilibili Transcript Scraper: 哔哩哔哩 Video to Text

Turn Bilibili (哔哩哔哩) videos into text, without an account or API key. Paste video links, b23.tv share links, the whole share text copied from the Bilibili app, BV or av IDs, or a creator's page, and each video comes back with its transcript in punctuated Chinese (as plain text and as timestamped segments), optional SRT and WebVTT subtitle files, and the video's metadata (title, creator, date, plays, likes, coins, favourites, danmaku, comments, shares, description, cover). Tick Translate to English to get an English translation next to the Chinese (paid option, see Pricing).

Multi-part videos (分P: courses, series) are read part by part, one record per part. Music-only videos come back as a free no_speech record instead of a made-up transcript. Deleted, private and members-only videos and failures are free too. You pay only for transcripts delivered.

What we measured

On the Apify platform on 16 October 2026, with the default settings (512 MB), on 20 inputs: video links, m.bilibili.com and b23.tv links, the share text of the Bilibili app, BV and av IDs, a creator's page, and damaged or unsupported links:

• 8 transcripts from 47 seconds to 1 minute 40: 5 in punctuated Chinese (Mandarin, and a Cantonese lesson), 2 in English, 1 song in Korean. 15 minutes billed in all.
• 3 videos without a transcript, free: a song with no clear words (no_speech), a music video whose language could not be settled (uncertain_language), a 74-minute video over the 30-minute default limit (too_long).
• 9 failures with their reason, free: a private video, a share link to a post, an expired share link, part 200 of a 155-part course, an anime episode, a live stream, an article, an ID with a missing character and a line of text.
• 0 failures on our side; 20 records in 45 seconds.

Longer videos come back complete too: a 14-minute and a 15-minute video of a science channel, about 4,500 characters each (15 October 2026, same settings).

Punctuated Chinese, complete, not guessed

Speech-to-text models write Chinese without punctuation unless they are guided: one block of characters with no commas and no full stops, unusable for subtitles, quotes or an LLM. We guide the model to punctuate, then put the punctuation in Chinese form (， 。 ？ ！ instead of , . ? !). Chinese text is joined without spaces, and wordCount counts each Chinese character as a word, and punctuation as none.

Fast talkers are transcribed in full. Speech models write only so much text for each 30 seconds of audio: very dense speech loses the end of each 30 seconds without a trace. We detect it and transcribe the same audio slowed down by 25% (same pitch), with the timestamps brought back to real time. If the text is still incomplete, a stronger model checks it; a video is never charged as a fragment.

Music is not speech. On songs, the model is known to invent sentences that were never said, often the credits of the videos it learnt from (「请不吝点赞 订阅 转发 打赏支持明镜与点点栏目」, 「字幕志愿者 …」). We filter those, along with phrases the model is not confident about and loops. A video whose speech is all filtered out is a free no_speech record; one where more than 20% of the speech had to be left out is a free incomplete_transcript record. Clearly sung lyrics still count as speech.

Speech-to-text runs on third-party AI providers: ElevenLabs Scribe v2, the most accurate model in our tests, for most videos (directly or through OpenRouter), and OpenAI's Whisper large-v3 turbo run by Groq for the first videos of some runs and as a backup. The audio is sent to them only to be transcribed. When we call ElevenLabs directly, it stores each transcript on its side: we delete it as soon as we have read it and check that it is gone; if a deletion cannot be confirmed before the run ends, the run log says so. Through OpenRouter, nothing is stored under our account and there is nothing for us to delete; OpenRouter's and ElevenLabs' own data policies apply.

Why not Bilibili's own subtitles? Bilibili shows its subtitles, including its AI subtitles, only to signed-in users, and we do not use accounts. Every transcript is made from the audio, so videos without subtitles are covered the same way.

What you can transcribe

Input	Examples	What you get
Videos	https://www.bilibili.com/video/BV1gHps6yEHL, m.bilibili.com/video/…	that video
IDs	BV1gHps6yEHL, av170001	that video
One part	https://www.bilibili.com/video/BV1Wv411h7kN?p=3	part 3 of a multi-part video
Share links	https://b23.tv/…, b23.tv/…	the video they lead to
Share text from the app	【影视飓风的直升机？怪东西分享9.0-哔哩哔哩】 https://b23.tv/…	the video in the text
Creators	https://space.bilibili.com/946974, uid:946974	their latest videos (maxVideosPerAccount, default 10); a skipped video is replaced by an older one

Mix them in one run. A video found in two inputs (a link and its share link, or a creator's page) is transcribed and charged once.

Multi-part videos. Some videos hold several parts (分P), each with its own title and length: a 155-part course, a series of episodes. Like on the site, a link without ?p= opens part 1 only; its record says how many parts there are (partCount) and how to get the others (otherParts). Add ?p=3 to the link for part 3, or tick All parts of multi-part videos (allParts: true) to transcribe every part, up to 50 per video: one record and one charge per part, with carouselIndex as the part number.

Videos only. Anime and series episodes (番剧, bangumi/play/ep…), live streams (live.bilibili.com), articles (read/cv…), posts (opus/…), search and channel pages get a free invalid_url or not_a_video record that says so. Members-only and paid videos, and videos Bilibili shows only to signed-in users, get a free unavailable record.

Damaged links are refused, never guessed. A link with a cut or extra character gets a free invalid_url record.

Languages

Most Bilibili videos are in Mandarin and come back in simplified Chinese (language: "zh"). Videos in another language (English, Japanese, Cantonese…) come back in that language: nothing is translated unless you tick Translate to English. We do not trust a single language detection when anything says otherwise: when the title's script and the detected language disagree, a stronger model settles the language and transcribes the video, as far as the run's own charges cover it. If it still cannot be settled, the video comes back as a free uncertain_language record, never as a charged guess. An English video with a Chinese title, alone in its run, can end this way: run it together with other videos.

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).

English translation (optional)

Tick Translate to English (translateToEnglish: true) and each transcript also comes back in English, in a translation field. The original text always stays in transcript, so you can check a sentence or quote it in Chinese. The translation is made from the whole punctuated transcript, sentence by sentence and in order, by a large language model; long talks are translated in full, never summarized.

• It is charged only when it is complete (translationStatus: "ok"). If it cannot be completed, you get the transcript as usual, translationStatus: "failed" with the reason in translationError, and nothing is charged for the translation.
• A transcript already in English is not translated and not charged (translationStatus: "notNeeded").
• Segments and subtitle files stay in the original language.
• Names of people, brands and shows are only as good as the speech-to-text heard them: when one matters, check it against the Chinese.

Output

One record per video (per part for a multi-part video). A transcript (status: "ok"), shortened here:

{
  "status": "ok",
  "platform": "bilibili",
  "videoId": "BV1rrJ66KExT",
  "url": "https://www.bilibili.com/video/BV1rrJ66KExT",
  "carouselIndex": null,
  "author": {"username": "万物60秒", "id": "3747570025236608", "url": "https://space.bilibili.com/3747570025236608"},
  "title": "一分钟走进梵高的一生",
  "partCount": 1,
  "publishedAt": "2026-10-15T01:53:30Z",
  "viewCount": 128,
  "durationSeconds": 61,
  "language": "zh",
  "transcript": "你知道吗，我画的一瓶向日葵后来一度是世界上最贵的画。可我活着的时候穷困潦倒。…如果能对我说一句话，你想说什么？",
  "segments": [
    {"start": 0, "end": 5.04, "duration": 5.04, "text": "你知道吗，我画的一瓶向日葵后来一度是世界上最贵的画。"},
    {"start": 5.68, "end": 8.1, "duration": 2.42, "text": "可我活着的时候穷困潦倒。"}
  ],
  "segmentCount": 13,
  "wordCount": 232,
  "billedMinutes": 2,
  "translation": "Did you know that a vase of sunflowers I painted later became the most expensive painting in the world? …",
  "translationStatus": "ok",
  "translatedMinutes": 2,
  "input": "https://www.bilibili.com/video/BV1rrJ66KExT",
  "inputType": "video"
}

A video without a transcript comes back as a free record that holds the reason and nothing else:

{
  "status": "error",
  "platform": "bilibili",
  "videoId": "BV1gHps6yEHM",
  "url": "https://www.bilibili.com/video/BV1gHps6yEHM",
  "errorCode": "private",
  "error": "this video is private (only its creator can see it)",
  "retryable": false,
  "input": "BV1gHps6yEHM",
  "inputType": "video"
}

Fields that need a word of explanation

• status: ok (transcript, charged), missing (the video exists but has nothing to transcribe: no_speech, incomplete_transcript, uncertain_language, no_audio, or too_long past your limit; free) or error (the link could not be read; free).
• errorCode and retryable: not_found (deleted video, expired share link, or a part number the video does not have), private, not_a_video (a share link that leads to a post or a creator) and invalid_url are final, and so is unavailable for a members-only or paid video. unavailable for a video under review, blocked, network, transcription_capacity and transcription_failed are worth retrying later (retryable: true). connection_failed means your own connection settings could not be used: fix them, or keep the default.
• segments: start, end and duration in seconds. Long Chinese segments are split at the end of sentences (at most 40 characters each, two subtitle lines); the time of each piece is shared out by its number of characters, so the bounds of the original segment are exact and the cuts inside it are close. includeSrt and includeVtt add ready-to-save .srt and .vtt files.
• carouselIndex, partTitle, partCount, otherParts: the part of a multi-part video (see above); carouselIndex is null for a single video.
• author.username is the creator's display name (unique on Bilibili, but the creator can change it); author.id is the UID, which never changes.
• viewCount, danmakuCount and the other counts are the video's, as Bilibili shows them when the record is made; for a multi-part video they are the totals of all its parts.
• filteredSegments: segments of the raw speech-to-text output left out as non-speech (an invented credit over music, a loop); minutes are billed on the video's length as usual.
• billedMinutes: started minutes of speech-to-text billed for this transcript.
• translation, translationStatus, translationError, translatedMinutes: only with Translate to English (see above). translatedMinutes is the number of started minutes charged for the translation.

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 run's summary is the OUTPUT record of the default key-value store: per input, transcripts, missing videos, errors and duplicates, plus speech-to-text statistics.

Input example

{
  "urls": [
    "https://www.bilibili.com/video/BV1gHps6yEHL",
    "https://www.bilibili.com/video/BV1Wv411h7kN?p=3",
    "https://space.bilibili.com/946974"
  ],
  "maxVideosPerAccount": 5,
  "includeSrt": true
}

Pricing

Pay per result: $0.001 per transcript plus $0.01 per started minute of speech, plus a start fee of $0.003 per GB of memory per run ($0.003 at the default 512 MB). A 1-minute video costs $0.011; a 14-minute video costs $0.141. Translate to English adds $0.03 per started minute of the video, only for complete translations: $0.561 for the same 14-minute video with its translation. Each part of a multi-part video (分P) is its own transcript. Videos without speech, videos over your length limit, private or deleted videos, errors and duplicates are free. Platform usage is included.

Your maximum charge per run is never exceeded: the length of each video is read before it is transcribed. A video that would cost more than what is left (with its translation, when you ask for one) gets a free over_budget record and the run goes on with the next ones (a shorter video may still fit); the run stops once not even one more transcript fits, and its status says both. 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 plus 2 per transcript have given no transcript, or once the audio transcribed for nothing (music-only videos) passes 15 minutes of audio plus the start fee and 40% of what the run has charged. A long video is never turned away because of its length: the run stops only once that much audio has really been transcribed for nothing. maxDurationMinutes (default 30, at most 60) skips longer videos before anything is transcribed.

Use cases

• Market and trend research in China: what creators in a niche say, searchable, quotable and ready for translation.
• Courses and lectures: a searchable, quotable text of each part of a Bilibili course (知识区), for notes or an LLM.
• Product and game research: what reviewers and players say about a product or a game, video by video.
• Subtitles: punctuated SRT files for your own Bilibili videos, for reposting elsewhere.
• LLM pipelines: summaries, translation and topic analysis across hundreds of Chinese videos.

Limits

• Videos only: no anime or series episodes (番剧), live streams, articles, posts, search or comments. For comments, danmaku and search, see our Bilibili Scraper.
• Private, members-only and paid videos come back as free private or unavailable records.
• 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. A video is charged only when its speech is clear; songs are charged only when the lyrics are clear on two passes that agree.
• Speed. Count about 10 to 30 seconds per video, more for long ones. Our speech-to-text capacity is shared fairly between runs; if it is temporarily used up, videos get free transcription_capacity records (retryable: true) and the status says when to retry.

Memory

512 MB is the default and the best setting: the run waits on speech-to-text, not on memory, and the Actor used at most 124 MB in our tests, two videos of 14 and 15 minutes included. The maximum is 1 GB.

FAQ

Is it legal to scrape Bilibili transcripts? This Actor reads only public videos, the way a visitor who is not signed in sees them, and accepts no Bilibili 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, PIPL in China). We keep nothing after the run. With Translate to English, the transcript text is sent to a language-model service that does not keep it to train its models.

Do I need a Bilibili account, a Chinese phone number or any setup? No. Keep the default connection: its cost is included in the price.

Can it translate? Yes, into English: tick Translate to English (charged only when the translation is complete). The Chinese transcript always comes with it.

What happens when Bilibili 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.

Is anything different on Apify's free plan? Runs started from an Apify free-plan account are transcribed by our Whisper models only, never by the ElevenLabs model we pay for per use, and the English translation option is off (nothing charged for it). Everything else is the same.

Does the Actor send anything to Quarra? Only when a run fails or runs out of time: the Actor, its build, the error type, the line of our code where it happened, and the memory and time the run was given, so we can fix it fast. Never your input, your results, your account or the run ID.