Audio & Video Transcription: Whisper, Podcasts, SRT/VTT

Turn audio and video files into text with OpenAI's Whisper speech recognition (open-source model, run on Apify's servers: no API key, no OpenAI account). For each file you get:

• Full transcript text and word count
• Timestamped segments (start, end, text): ready for search, chaptering or quoting
• SRT and VTT subtitle files (download links)
• Language detected automatically (90+ languages) with a confidence score, or set it yourself
• Translate to English: English text from speech in any language (Whisper's built-in translation)
• Word-level timestamps (optional): start, end and confidence for every word, for captions and editing
• Duration and the minutes billed

Inputs it accepts:

• Direct links to audio or video files: mp3, m4a, aac, wav, flac, ogg, opus, mp4, mov, webm, mkv…
• Podcast RSS feeds: transcribes the latest episodes
• Apple Podcasts links: finds the show's feed and transcribes the latest episodes

Silent files, files with no speech, unreachable links and files that can't be decoded are not charged.

Who it's for

• Podcasters and creators: show notes, subtitles, blog posts and quotes from episodes.
• Researchers and journalists: searchable text of interviews, lectures and recordings.
• AI and RAG pipelines: feed spoken content into LLMs, summaries and search.
• Accessibility: captions for your own videos (SRT/VTT).

Input

Option	Default	What it does
Audio/video files or podcasts	–	File links, RSS feeds or Apple Podcasts links
Accuracy	Balanced	Fast (Whisper tiny), Balanced (base), Accurate (small)
Language	auto	Two-letter code to skip detection, e.g. en, ja
Translate to English	off	English output from any language
Word-level timestamps	off	Per-word timing and confidence
Podcasts: episodes per feed	1	Latest N episodes per feed
Max minutes per file	180	Longer files are skipped (not charged)
Save SRT and VTT subtitle files	on	Links in each result

Output

A real result (LibriVox public-domain recording of the Gettysburg Address):

{
  "url": "https://archive.org/download/gettysburg_address_librivox/gettysburg_address_lincoln_64kb.mp3",
  "status": "ok",
  "language": "en",
  "languageProbability": 0.998,
  "durationSec": 161,
  "billedMinutes": 3,
  "model": "base",
  "processingSec": 36.6,
  "wordCount": 328,
  "text": "This is a Libravox Recording. All Libravox recordings are in the public domain. ... The address set the dedication of the National Cemetery at Gettysburg, November 19, 1863, ...",
  "segments": [{ "start": 0.0, "end": 4.5, "text": "This is a Libravox Recording." }],
  "srtUrl": "https://api.apify.com/v2/key-value-stores/.../records/transcript-001.srt",
  "vttUrl": "https://api.apify.com/v2/key-value-stores/.../records/transcript-001.vtt"
}

Note the transcript says "Libravox": Whisper writes what it hears, and names can come out slightly wrong. The Accurate setting helps.

Other statuses: unreachable, no_speech, too_long, not_supported, over_budget, error. None are charged.

Use it from your AI agent (MCP)

Claude, Cursor and other MCP clients can call this tool directly through Apify's MCP server. Add it as an MCP server / connector:

{
  "mcpServers": {
    "apify": {
      "url": "https://mcp.apify.com?tools=swiftkit/transcribe",
      "headers": { "Authorization": "Bearer <YOUR_APIFY_TOKEN>" }
    }
  }
}

Then just ask, for example: "Transcribe this episode and give me the five main points: https://…/episode.mp3". Each call is billed like a normal run.

Pricing

• A small fee per run (loading the speech model)
• Per audio minute transcribed, rounded up per file. Fast and Balanced cost the same. Accurate needs about 3x the computing power, so each minute counts as 3.

See the Pricing tab. With 2 CPU cores (the default 8 GB memory), Balanced transcribes about 8x faster than real time: a 60-minute episode takes about 7–8 minutes.

Limits, honestly

• Only files you're allowed to use. Transcribe your own recordings, public-domain or licensed material, or podcasts for personal or research use. Respect copyright.
• Video and social platforms aren't supported (YouTube, TikTok, Instagram, Facebook, X, Vimeo, Spotify…). Their terms forbid downloading. Give a direct file link instead.
• No speaker labels (who said what). Translation is into English only. Whisper is very good, not perfect: check names, numbers and technical terms.
• Files up to 1.5 GB and 10 hours.

More tools from SwiftKit

• Podcast Search & Episodes: find shows, episode lists with audio links, top charts
• PDF to Text & Markdown: text from PDFs for AI and RAG
• Website to Markdown for AI: clean page content for LLMs

Questions?

Open an issue on the Issues tab.