Bilibili Video Transcriber – B站视频转文字

Turn Bilibili (B站) videos into text. Paste video links or BV ids and get a clean transcript with sentence-level timestamps, ready-made SRT / WebVTT subtitles, an AI English summary of every video, an optional full English translation, and the video's metadata (title, creator, stats, publish time). Works for Mandarin, Cantonese, English, Japanese and Korean.

• Batch friendly. Process one video or hundreds in one run, including multi-part (分P) videos.
• Cheap and predictable: $0.005 per audio minute. A 10-minute video costs about $0.05. Failed or skipped videos cost nothing.
• Public videos only. Paid and members-only content (charge-exclusive 充电专属 videos, paid courses 课堂, bangumi/films) is detected and skipped for free, before any audio is downloaded. See Supported content.
• English summary included (new in 1.1). Every transcribed part gets a short English bullet-point summary at no extra cost. Turn it off with englishSummary: false.
• Optional full English translation (new in 1.1). Set translateToEnglish: true to get the whole transcript in English, for $0.002 per audio minute, charged only if the translation succeeds.
• Chinese speech recognition. Uses the open-source SenseVoice-Small speech model, which supports Mandarin, Chinese-English code-switching, Cantonese, English, Japanese and Korean. Punctuation is added automatically. Transcripts are machine-generated and can contain mistakes.

Why use it?

Many Bilibili videos have no subtitles. This Actor recognizes the audio directly, so you get text even for videos that never had subtitles.

Typical uses:

• Build RAG / LLM knowledge bases from Chinese video content
• Read Chinese videos in English: skim the English summary, or get the full translation
• Summarise, translate or search Bilibili videos
• Accessibility: subtitles and transcripts for deaf and hard-of-hearing viewers
• Personal study and research: language learning, note-taking, academic and content research
• Help non-Chinese speakers understand and research Bilibili content
• Do market and content research on Chinese creators, products and trends

Why not just use a free tool?

Free scripts and desktop apps can transcribe Bilibili videos too, but you have to run and maintain them yourself. With this Actor:

• Nothing to install. No speech model, ffmpeg or GPU on your own computer. Everything runs in Apify's cloud.
• Batch in the cloud. Send one video or hundreds in a single run, and let it work while your computer is off. Schedule it or trigger it from your own code.
• Callable by AI agents. Use it as an MCP tool in Claude, Cursor and other MCP clients, or call it with one REST/Python request (see Use from AI agents).
• Structured, ready-to-use output. JSON with sentence timestamps, SRT/WebVTT subtitles, video metadata and an English summary in the same item, with no post-processing needed.
• Pay only for what you use. $0.005 per audio minute and no subscription. Failed and skipped videos are free.

Supported content

The Actor only transcribes public Bilibili videos that anyone can watch without logging in or paying. These inputs are detected and skipped. You get a "type": "skipped" item with the reason, you aren't charged, and no audio is downloaded:

Skipped type	How it's detected	errorCode
Charge-exclusive videos (UP主充电专属, including 单片付费 and 抢先看 while it's still exclusive)	is_upower_exclusive / is_upower_exclusive_with_qa in the public video info	UPOWER_EXCLUSIVE
Paid courses (B站课堂, bilibili.com/cheese/... links or b23.tv links to them)	Link type, or a redirect_url pointing to /cheese/	PAID_COURSE
Paid videos and paid collections (UGC付费, 付费合集)	rights.ugc_pay / rights.arc_pay, or a paid collection with no public audio	PAID_CONTENT
Videos where only a short preview is public	The public stream is much shorter than the video	PREVIEW_ONLY
Bangumi, films and TV (番剧/影视, /bangumi/ links, PGC paid content)	Link type, redirect_url or rights.pay	UNSUPPORTED

Private, deleted, under-review and region-locked videos produce an error item (NOT_FOUND, UNDER_REVIEW, RESTRICTED, NO_AUDIO), which is also free. Course episodes entered by their av id usually aren't visible through the public video API, so they come back as NOT_FOUND.

Input

Field	Description	Default
videos	Video URLs, BV ids (BV1Eapz6PErV), av ids (av170001) or b23.tv short links. Add ?p=N to a URL to choose a part.	–
parts	first (the first part, or the ?p= part) or all parts of multi-part videos	first
language	zh (Mandarin, recommended), auto, en, yue, ja, ko. With zh, parts that are clearly English, Japanese, Korean or Cantonese are detected and switched automatically.	zh
outputFormats	Any of text, segments, srt, vtt. Add srt to get SRT subtitles built from the speech-recognition segment timestamps (off by default, no extra charge).	text, segments
saveSubtitleFiles	Also save <BVID>_p<N>.srt / .txt files to the key-value store	false
englishSummary	Add an AI-written English summary (up to 5 bullet points) per part. Included in the price.	true
translateToEnglish	Add a full English translation of the transcript per part. Extra charge: $0.002 per audio minute, only when it succeeds.	false
maxLlmCharsPerVideo	Safety cap on transcript characters sent to the AI per part (max 30000 ≈ 60–75 min of Chinese speech). Longer text is cut and flagged LLM_INPUT_TRUNCATED.	30000
qualityCheck	Detect the spoken language and add dialect / accent / noise warnings (no extra charge)	true
includeMetadata	Include the description, stats, category, publish time and cover	true
maxVideos	Maximum number of parts per run	50
maxAudioMinutesPerVideo	Skip parts longer than this (they aren't charged)	120

Example:

{
  "videos": ["https://www.bilibili.com/video/BV1Eapz6PErV", "BV15Lpf62EG8"],
  "outputFormats": ["text", "segments", "srt"]
}

With a full English translation:

{
  "videos": ["BV1sapm6KE2q"],
  "translateToEnglish": true
}

Output

You get one dataset item per video part:

{
  "type": "transcript",
  "input": "https://www.bilibili.com/video/BV1Eapz6PErV",
  "bvid": "BV1Eapz6PErV", "aid": 117434522538286, "cid": 42731438751, "page": 1, "partsCount": 1,
  "url": "https://www.bilibili.com/video/BV1Eapz6PErV",
  "title": "原来，小猫是这样认识武警哥哥的～",
  "durationSec": 60,
  "publishedAt": "2026-10-14T07:00:00+08:00",
  "creator": { "mid": 456664753, "name": "央视新闻" },
  "stats": { "view": 1631199, "like": 208454, "coin": 3893, "favorite": 15862, "share": 2614, "danmaku": 635, "reply": 2197 },
  "transcript": {
    "engine": "sensevoice-small-int8@sherpa-onnx-1.13.8",
    "requestedLanguage": "zh",
    "language": "zh",
    "text": "当时就觉得有一个毛茸茸的东西从我的脚边来回穿梭。……",
    "segments": [
      { "start": 0.25, "end": 5.63, "text": "当时就觉得有一个毛茸茸的东西从我的脚边来回穿梭。" },
      { "start": 5.63, "end": 11.03, "text": "根据稍微纪律来讲的话，我要平视前方，……" }
    ],
    "charCount": 275, "segmentCount": 15, "speechSec": 55.0, "audioSec": 60.0, "flags": [],
    "quality": { "confidence": "high", "probeDisagreement": 0.0, "probeSampleSec": 35.8, "probeLanguages": { "zh": 1.0 }, "charsPerSpeechSec": 4.69, "speechRatio": 0.879 }
  },
  "englishSummary": {
    "status": "OK",
    "text": "- A guard noticed something furry moving around his feet; though discipline required looking straight ahead, he checked and was surprised and delighted to see a kitten.\n- One rainy day, the kitten took shelter on his post, and with few people around he stepped over slightly to shield it from the wind and rain.\n- Over time, the kitten and the guard grew closer, and the cat often stayed by his side while he was on duty.\n- He reflected that cats are very alert animals and may seek out someone who makes them feel safe.\n- He said that as a people's armed police officer, providing a sense of security to the people is part of the job, which may be why the kitten came to them.",
    "model": "deepseek-v4.1-flash", "inputChars": 275, "totalChars": 275, "inputTruncated": false,
    "latencySec": 2.56, "promptTokens": 283, "completionTokens": 143
  },
  "billedAudioMinutes": 1,
  "scrapedAt": "2026-10-15T06:40:37+00:00"
}

With translateToEnglish: true the item also has:

"englishTranslation": {
  "status": "OK",
  "text": "Recently, a project called RAEA suddenly became popular for no apparent reason. Looking at the official website, it's all pretentious, and the title clearly looks like it was made by AI. Its full name is very domineering: reverse engineer anything. …",
  "model": "deepseek-v4.1-flash", "inputChars": 2848, "totalChars": 2848, "inputTruncated": false, "chunks": 2,
  "latencySec": 12.76, "residualChineseRatio": 0.0, "billedTranslationMinutes": 7
}

English summary and translation

• Written by an LLM (DeepSeek, via Alibaba Cloud Model Studio) from the machine transcript, so ASR mistakes can carry over. Treat them as a reading aid.
• status is OK, FAILED or SKIPPED. A failed summary or translation never affects the transcript: the transcript is returned as usual, and transcript.flags gets SUMMARY_FAILED or TRANSLATION_FAILED, with the reason in errorCode / message. Failed translations are not charged.
• SKIPPED reasons: NO_SPEECH (nothing to summarize), ALREADY_ENGLISH (English parts aren't translated or charged), CHARGE_LIMIT_REACHED (the translation would go over your max cost per run).
• Long transcripts are translated in chunks of about 2000 characters. At most maxLlmCharsPerVideo characters (default and max 30000) are sent per part; longer text is cut, flagged LLM_INPUT_TRUNCATED, and only the translated share of the minutes is charged.
• Skipped (paid/members-only) and failed videos are never sent to the AI.
• residualChineseRatio is the share of Chinese characters left in the translation (normally 0).

Quality warnings

transcript.quality.confidence is high, medium or low. transcript.flags lists warnings. The flags come from heuristics: treat them as hints, not guarantees.

Flag	Meaning
POSSIBLE_DIALECT_ACCENT_OR_NOISE	Dialect, a strong accent or noisy/outdoor audio is likely, so expect more errors
LOW_CONFIDENCE	The transcript is probably unreliable (strong dialect, wind noise, singing)
LANGUAGE_SWITCHED	You asked for zh, but the part was clearly another supported language, so it was transcribed in that language (see transcript.language)
MIXED_OR_UNSUPPORTED_LANGUAGE	No supported language dominates (for example Spanish), so the text is unreliable
CANTONESE_DETECTED	Cantonese speech was detected
FOREIGN_SCRIPT_IN_TEXT	Japanese or Korean characters appeared in a Chinese transcript
LOW_SPEECH_RATIO / NO_SPEECH_DETECTED	Little or no speech (music, silent screen recordings)
SUMMARY_FAILED / TRANSLATION_FAILED	The English summary / translation failed (the transcript is fine). See englishSummary / englishTranslation for the reason.
LLM_INPUT_TRUNCATED	The transcript was longer than maxLlmCharsPerVideo; the summary/translation covers only the first part

How the check works: a sample of about 20–90 s of speech is decoded a second time with automatic language detection. Two signals are used: how much the two decodes disagree, and how many characters come out per second of speech (clean Mandarin is about 5 per second). Both were calibrated on 18 test videos. Clean news and narration were never flagged. All dialect videos, wind noise, crowd noise and concerts were flagged. An outdoor street-food video was flagged medium. Kitchen noise wasn't flagged, because the model coped well with it. There is no per-word confidence score.

Inputs that fail produce an error item instead, and you aren't charged for them:

{ "type": "error", "input": "BV1xxxxxxxxx", "errorCode": "NOT_FOUND", "message": "..." }

Paid or unsupported inputs produce a free skipped item:

{ "type": "skipped", "input": "BV1oMBUBREK8", "errorCode": "UPOWER_EXCLUSIVE", "bvid": "BV1oMBUBREK8", "billedAudioMinutes": 0, "message": "charge-exclusive video (UP主充电专属): ..." }

Error code	Meaning
INVALID_INPUT	No BV/av id found in the input
NOT_FOUND / UNDER_REVIEW	The video was deleted, is private or is under review
RESTRICTED / NO_AUDIO	Region-locked or members-only content, or no audio stream
UPOWER_EXCLUSIVE / PAID_COURSE / PAID_CONTENT / PREVIEW_ONLY / UNSUPPORTED	Paid or unsupported content. These come as "type": "skipped" items; see Supported content
TOO_LONG	The part is longer than maxAudioMinutesPerVideo
BILI_RISK_CONTROL	Bilibili rate-limited the request. Please retry later.
DOWNLOAD_FAILED / ASR_FAILED	Temporary failure. Retry.
CHARGE_LIMIT_REACHED	Your max cost per run was reached, so the run stopped

Use from AI agents (MCP) and the API

MCP (Claude, Cursor, any MCP client): add the Apify MCP server with only this Actor loaded:

{
  "mcpServers": {
    "apify": { "url": "https://mcp.apify.com?tools=lone_fascia/bilibili-video-transcriber" }
  }
}

On first connection your browser opens to sign in to Apify (or send Authorization: Bearer <APIFY_TOKEN>). Then just ask your agent, e.g. "Transcribe https://www.bilibili.com/video/BV1xx411c7mD and summarize it." The agent calls the Actor and reads the results with get-dataset-items.

REST API (one call, returns the transcript items):

curl -X POST "https://api.apify.com/v2/acts/lone_fascia~bilibili-video-transcriber/run-sync-get-dataset-items" \
  -H "Authorization: Bearer $APIFY_TOKEN" -H "Content-Type: application/json" \
  -d '{"videos": ["https://www.bilibili.com/video/BV1xx411c7mD"], "outputFormats": ["text", "segments", "srt"]}'

Python:

from apify_client import ApifyClient
client = ApifyClient("<APIFY_TOKEN>")
run = client.actor("lone_fascia/bilibili-video-transcriber").call(
    run_input={"videos": ["BV1xx411c7mD"], "outputFormats": ["text", "srt"]})
for item in client.dataset(run["defaultDatasetId"]).iterate_items():
    print(item.get("transcript", {}).get("text"))

Pricing

Pay per event: $0.005 per audio minute (event audio-minute), English summary included. Optional full English translation: +$0.002 per audio minute (event translation-minute), only for parts whose translation succeeded.

• You're charged per started minute of each successfully transcribed part: a 61-second video counts as 2 minutes.
• Parts where no speech is detected are charged the 1-minute minimum.
• Failed, skipped (paid/unsupported) and too-long parts are free.
• SRT/VTT subtitles, quality checks, metadata and the English summary don't cost extra.
• Translation is charged per started audio minute of each successfully translated part. Failed translations, English parts and skipped parts are free.
• The Actor respects your Max cost per run setting and stops before it would go over.

Video length	Transcript + English summary	With full English translation
1 min	$0.005	$0.007
10 min	$0.05	$0.07
1 hour	$0.30	$0.42

Speed and limits

• Recommended memory: 4 GB (the default and the maximum). One transcription thread, with peak memory around 0.9 GB. A 10-minute video takes about 1.5 minutes.
• 2 GB also works, but it's about 2× slower: a 10-minute video takes about 3–4 minutes. The price per audio minute is the same.
• The English summary adds about 3–6 seconds per part. A full translation adds about 2–3 seconds per audio minute (a 10-minute video: about 20–40 seconds).
• Each part can be up to 300 minutes long. The default skip limit is 120 minutes.
• Requests to Bilibili are spaced out on purpose, to be polite and to avoid rate limits.

Compliance, privacy and fair use

• Public videos only. The Actor only processes public videos that anyone can watch without logging in or paying. Paid content and charge-exclusive (充电专属) videos are detected and skipped.
• No danmaku, comments or user data. The Actor does not collect danmaku (bullet comments), comments, or any viewer/user personal data. It reads only public video information (title, creator name, public stats) and the lowest-bitrate audio stream needed for transcription.
• Audio is deleted after processing. Audio is deleted right after it's transcribed; no audio or video is stored or returned.
• Third-party AI processing. When englishSummary or translateToEnglish is on, the transcript text (not the audio) is sent to Alibaba Cloud Model Studio (阿里云百炼, servers in mainland China) for processing with a DeepSeek model. If you don't want that, set englishSummary: false (and leave translateToEnglish off).
• Not affiliated with Bilibili. This Actor is an independent tool and is not affiliated with, endorsed or sponsored by Bilibili. Bilibili is a trademark of its respective owner.
• Transcripts are machine-generated and can contain mistakes, especially with strong dialects, heavy background music or overlapping speakers.
• You're responsible for how you use the transcripts. Respect the creators' copyright and Bilibili's terms of service, and don't republish content you don't have the rights to.

---

中文说明

B站视频转文字：把 B站视频链接或 BV 号贴进来，就能拿到带时间戳的逐句文字稿、可以直接用的 SRT / VTT 字幕、AI 英文摘要、可选的全文英文翻译，以及视频元数据（标题、UP主、播放/点赞/投币/收藏等数据、发布时间）。

• 很多视频没有字幕，本工具直接识别音频，没有字幕的视频也能转写。
• 支持普通话、粤语、英语、日语、韩语，中英夹杂也能识别，自动加标点。language 默认是 zh（普通话）：自动识别模式在口音重的普通话里容易混进日文和韩文字符，所以默认固定为中文。如果某个视频明显是英语、日语、韩语或粤语，会自动切换成对应语言，并标记 LANGUAGE_SWITCHED。
• 和免费工具比：不用在自己电脑上装语音模型、ffmpeg 或显卡，全部在 Apify 云端运行；一次可以批量跑几百个视频，关机也不影响；AI agent 可以通过 MCP 或 API 直接调用；输出是现成的结构化 JSON（时间戳、字幕、元数据、英文摘要），按用量付费，没有订阅费。
• 支持批量处理和多P视频（parts 选 all 就会转写所有分P）。
• 价格：每音频分钟 $0.005，10 分钟的视频约 $0.05，已包含英文摘要。失败、跳过、超长的视频不收费；没有检测到人声的视频按最低 1 分钟计费。
• 英文摘要（1.1 新增，默认开启）：每个转写成功的分P都会附带一段英文要点摘要（最多 5 条，字段 englishSummary.text），由大模型（阿里云百炼 DeepSeek）根据文字稿生成，不额外收费；不需要可以设 englishSummary: false。
• 全文英文翻译（1.1 新增，可选）：设 translateToEnglish: true 后输出整篇英文译文（字段 englishTranslation.text），长文本自动分块翻译。按音频分钟另收 $0.002（收费事件 translation-minute），只在翻译成功后收费；翻译失败、英文视频、跳过的视频都不收费。10 分钟视频含翻译约 $0.07。
• AI 失败不影响文字稿：摘要或翻译失败时，文字稿照常返回，transcript.flags 里会标 SUMMARY_FAILED / TRANSLATION_FAILED，原因写在 englishSummary / englishTranslation 的 errorCode、message 里。每个分P最多送 maxLlmCharsPerVideo（默认也是上限 30000）字给大模型，超出部分截断并标记 LLM_INPUT_TRUNCATED，翻译只按实际翻译的比例收费。付费/充电等被跳过的视频不会调用大模型。
• 摘要和翻译基于机器转写结果，转写错误可能被带入，仅供参考。开启摘要/翻译时，文字稿（不含音频）会发送到阿里云百炼（服务器在中国大陆）处理；不需要可关闭 englishSummary（并保持 translateToEnglish 关闭）。
• 只支持公开视频（不登录、不付费就能看的视频）。下面这些类型会被识别出来并直接跳过：输出一条 "type": "skipped" 记录并写明原因，不收费，也不下载音频：
  • UP主充电专属视频（包括单片付费、仍在抢先看期内的视频）→ UPOWER_EXCLUSIVE
  • B站课堂付费课程（bilibili.com/cheese/... 链接，或者跳转到课堂的 b23.tv 短链）→ PAID_COURSE
  • 付费视频 / 付费合集（UGC付费）→ PAID_CONTENT；只公开了试看片段的视频 → PREVIEW_ONLY
  • 番剧、电影、电视剧（/bangumi/ 链接、PGC 付费内容）→ UNSUPPORTED
  • 私密、已删除、审核中或有地区限制的视频会返回错误记录（NOT_FOUND / UNDER_REVIEW / RESTRICTED / NO_AUDIO），同样不收费。
• SRT 字幕：在 outputFormats 里勾选 srt（默认不开），就会按语音识别的分段时间戳生成标准 SRT 字幕；勾选 vtt 生成 WebVTT。不额外收费。
• 输入支持：完整链接、BV 号、av 号、b23.tv 短链；链接后面加 ?p=N 可以指定分P。
• 输出：每个分P一条数据。transcript.text 是全文，transcript.segments 是逐句时间戳，srt / vtt 是字幕文本。开启 saveSubtitleFiles 后，还会把 .srt / .txt 文件保存到 Key-value store，可以直接下载。
• 给 AI agent / MCP 用：在 MCP 客户端（Claude、Cursor 等）里添加 https://mcp.apify.com?tools=lone_fascia/bilibili-video-transcriber，登录 Apify 后直接对 agent 说"把这个 B站视频转成文字并总结"即可；也可以用 REST API run-sync-get-dataset-items 或 Python apify-client 一次调用拿到结果（示例见上方英文部分）。
• 适用场景：无障碍字幕（方便听障用户）、个人学习与研究、搭建 AI 知识库（RAG）、视频总结、帮海外用户看懂和研究 B站内容。
• 质量提示：transcript.quality.confidence 取 high / medium / low 三档；transcript.flags 会标出 POSSIBLE_DIALECT_ACCENT_OR_NOISE（可能是方言、口音重或户外噪声）、LOW_CONFIDENCE（结果可能不可靠）、MIXED_OR_UNSUPPORTED_LANGUAGE（不支持的语言或多种语言混杂）、CANTONESE_DETECTED（检测到粤语）、NO_SPEECH_DETECTED / LOW_SPEECH_RATIO（没有或很少人声）等。这些标记是启发式判断，只作参考。
• 内存：推荐默认的 4 GB（也是上限），10 分钟的视频大约 1.5 分钟处理完；2 GB 也能用，但慢一倍左右（大约 3–4 分钟），每分钟价格不变。
• 说明：识别结果由机器生成，方言、背景音乐很重或多人同时说话时可能有错误。请遵守版权和 B站用户协议，不要未经授权转载他人内容。

合规与隐私

• 只处理公开视频：只处理不登录、不付费就能看的公开视频；付费内容和充电专属视频会被识别并跳过。
• 不抓弹幕和用户数据：不采集弹幕、评论或任何观众/用户个人数据，只读取公开的视频信息和转写所需的最低码率音频。
• 音频处理完即删：音频转写完成后立即删除，不保存、不返回音频或视频。
• 第三方 AI 处理：开启摘要/翻译时，文字稿会发送到阿里云百炼（服务器在中国大陆）处理；不需要可关闭 englishSummary。
• 声明：本工具与哔哩哔哩无关联，Bilibili 是其所有者的商标。