Whisper Transcriber — audio & video to text, SRT & VTT

Send audio or video URLs — or a podcast RSS feed — and get back the transcript: full text, timestamped segments, and ready-to-save SRT + VTT subtitle files. Whisper runs inside the Actor (faster-whisper, int8, CPU), so there is no API key, no external service, and no per-provider markup. 99 languages, auto-detected. The default tiny model is fast, draft-quality transcription that does well on clear English speech; pick base for other languages, accents or noisy audio. Optional translation to English.

Built on the open-source Whisper model (MIT licence); not affiliated with or endorsed by OpenAI.

Built for AI agents, podcast pipelines, subtitle workflows, meeting-notes automation and anyone who needs speech as text without wiring up a transcription provider.

What you get

One dataset item per media file:

{
  "url": "https://example.com/episode-42.mp3",
  "episode_title": "Episode 42 — Pricing",
  "status": "ok",
  "model": "tiny",
  "language": "en",
  "language_probability": 0.994,
  "duration_seconds": 1864.3,
  "billed_minutes": 32,
  "n_segments": 214,
  "n_words": 4930,
  "text": "Full transcript as one string …",
  "segments": [
    { "start": 0.0, "end": 6.48, "text": "Welcome back to the show." }
  ],
  "srt": "1\n00:00:00,000 --> 00:00:06,480\nWelcome back to the show.\n",
  "vtt": "WEBVTT\n\n00:00:00.000 --> 00:00:06.480\nWelcome back to the show.\n"
}

Sources

• Direct media files — mp3, m4a, aac, ogg, opus, wav, flac, mp4, webm and anything else FFmpeg can decode. Up to 250 MB per file.
• Podcast RSS/Atom feeds — the feed URL expands to its newest episodes (Episodes per RSS feed, default 1, up to 10).

Not supported, deliberately: downloading from YouTube, TikTok or other social platforms. Send direct media files you have the right to process.

Pricing — pay only for delivered transcripts

Event	Price	Charged when
file-transcribed	$0.005	a file delivered a non-empty transcript
transcription-minute	$0.008	per started minute of audio transcribed

A 30-minute podcast episode costs $0.245. A one-hour meeting costs $0.485. Compare: hosted transcription APIs charge $0.36–0.90 per hour plus you build the pipeline.

The billing rule is deliberate: files that fail to download, cannot be decoded, or contain no recognizable speech are never billed — not even the per-file fee. If you got no transcript, you pay nothing. Your cost ceiling per file = $0.005 + $0.008 × Max minutes per file (default 120).

Options

Field	Default	Notes
media_urls	—	file URLs and/or RSS feed URLs, up to 25
model	tiny	tiny (fastest; clear English) or base (more accurate, ~2–3× slower)
language	auto	ISO code (en, ar, es, …) skips detection
task	transcribe	translate = English output from any language
rss_episodes	1	newest N episodes per feed
max_minutes_per_file	120	hard cost/length cap; longer files flagged truncated
vad_filter	true	skip silence — faster, fewer hallucinated captions
word_timestamps	false	per-word timing inside segments (slower)

Both models are built into the Actor image and start instantly — nothing is downloaded at run start (the Actor runs with the model hub switched off, so it never checks for or fetches model updates). The price per minute is the same for both; base only makes the run take longer.

Change, September 2026: the small model was retired (it decoded about 3× slower than base, too slow for the per-minute price) and is no longer accepted by the input schema — a run that asks for small is rejected before it starts. Switch existing tasks and integrations to base.

Notes on quality

• Voice-activity detection is on by default — silent stretches are skipped instead of becoming hallucinated captions, and long quiet files transcribe much faster.
• tiny is fast, draft-quality transcription: fine for clear English speech. For other languages, accents or noisy audio choose base — the more accurate of the two models (Whisper's own benchmarks put tiny's error rate well above base's, and the gap widens outside English).
• Set language explicitly for very short clips; auto-detection has less signal to work with under ~10 seconds.
• translate produces English text directly — no second pass needed.

Limits, stated plainly

• CPU transcription takes real time. As a rough guide on the default 4 GB run, budget 15–30 seconds of processing per audio minute with tiny and about 2–3× that with base (speed depends on the audio and the model — these are guide figures, not a guarantee). The default run timeout is 60 minutes — raise it for multi-hour files, and for anything over about an hour with base.
• 250 MB per file, 25 URLs per run, 480 minutes max per file.
• Live streams are not supported — the URL must be a finite file.

Data handling

• Your media is fetched from the URLs you provide into the run's own container, transcribed there, and the temporary file is deleted as soon as that file is done. Inputs and outputs are processed only inside the Apify platform, for the duration of the run.
• No third-party service and no sub-processor beyond the Apify platform itself: Whisper runs locally in the container (the model weights ship in the image and the model hub is switched off at run time), so no audio, transcript or metadata is sent to OpenAI or to anyone else. The only outbound requests are to your media/feed URLs and to the Apify API that stores your run's results.
• Transcripts land in your run's default dataset and are kept under your Apify plan's normal storage retention; the Actor keeps nothing of its own beyond the run.

The family

Actor	Use it for
Whisper Transcriber (this one)	speech in audio/video → text + subtitles
PDF Text Extractor	fast clean text from digital PDFs
PDF OCR Extractor	scanned PDFs, Arabic + English OCR
Doc to Markdown	documents into LLM-ready Markdown

Issues and requests

Open an issue on the actor's Issues tab — I read all of them.