Audio & Video Transcription (offline Whisper)

Turn recordings into text: interviews, meetings, lectures, voice notes, podcast episodes you have as files, and the sound track of videos. Give it links to audio or video files, or send files inline as base64, and get back for each one: the full text, timed segments, the detected language, and SubRip (.srt) and WebVTT (.vtt) subtitle files.

Everything runs inside the Actor with OpenAI's Whisper speech model (base, via faster-whisper). Your files are not sent to any third-party API.

See it run: a 90-second video shows this page, a sample clip with the subtitles one real run returned, and a Python example that calls the Actor from your own code. We made the video; its narration is a synthetic voice.

Podcasts: a second example reads a podcast's RSS feed with our RSS Feed Reader, transcribes only the episodes it has not done yet, and searches every transcript for the minute something was said.

What you get

• Text of each file, and segments with start and end times in seconds (includeSegments, on by default).
• Word timestamps (wordTimestamps): each word's start, end and probability, for karaoke-style captions or search.
• Subtitles: .srt and .vtt files saved in the run's key-value store, linked from each item (subtitles), or embedded in the item too (subtitlesInItem).
• 100 languages, detected automatically (language and languageProbability) or set by you; or translation of the speech to English text (task: translate).
• Audio and video files: MP3, M4A/AAC, MP4, MOV, 3GP, WAV, FLAC, ALAC, OGG/Vorbis, Opus, WebM, MKV, WMA/ASF, AVI, FLV, MPEG-TS, AIFF, CAF, AMR, AC-3/E-AC-3, DTS, WavPack, APE and more. Only the first audio track is used; video pictures are never decoded.
• Silence and music skipped before transcription (vadFilter, on by default, Silero VAD), which cuts made-up text in long pauses. Timestamps still refer to the original file.
• Long files are transcribed in 10-minute pieces split at a pause, so memory stays flat; the language found in the first piece is kept for the rest.

Failed files are reported in the dataset with a reason and are not charged: fetch-failed, file-too-large, robots-disallowed, unsupported-file, no-audio-track, unsupported-codec, undecodable-audio, invalid-base64, time-limit.

Respectful by design

• Only the files you name are downloaded, one request each. The site's robots.txt is read first and obeyed, and downloads from one site are spaced (500 ms by default). No proxy, no browser, no login.
• Only public internet addresses are fetched: a URL or redirect into a private, loopback or link-local network is refused.
• Links must point to the media file itself. Pages on video platforms (YouTube, TikTok, Instagram) are not files, and this Actor does not extract media from them.

Use it responsibly

From the Whisper model card:

• Consent. Transcribe only recordings you have the right to process. OpenAI cautions "against using Whisper models to transcribe recordings of individuals taken without their consent".
• Made-up text. The model can output text that was never spoken ("hallucination"), most often in silence, music or noise, and can repeat itself. Check transcripts that matter against the audio.
• Not for high-risk decisions. Do not use the output on its own for decisions about people (legal, medical, employment, credit and the like). OpenAI recommends "against use in high-risk domains like decision-making contexts".
• Accuracy varies by language, accent, recording quality and background noise. base is a small, fast model: it is good on clear speech in widely spoken languages and weaker on others. Translation to English is rougher than transcription.

Input

Give fileUrls, filesBase64 or both. Results come back in input order. Apify limits the whole input to a few MB, so send larger files as links. The default run memory is 1 GB, which suits files up to about 300 MB; give larger files 2 GB or more.

Output

One dataset item per file (this is the Store example, a 24-second spoken notice):

{
  "source": "https://example.com/audio/speech-sample.mp3",
  "url": "https://example.com/audio/speech-sample.mp3",
  "finalUrl": "https://example.com/audio/speech-sample.mp3",
  "fileName": "speech-sample.mp3",
  "status": "ok",
  "error": null,
  "container": "mp3",
  "audioCodec": "mp3float",
  "hasVideo": false,
  "durationSeconds": 24.005,
  "transcribedSeconds": 24.005,
  "speechSeconds": 24.005,
  "billedMinutes": 1,
  "truncated": false,
  "language": "en",
  "languageProbability": 0.998,
  "languageDetected": true,
  "task": "transcribe",
  "model": "base",
  "segmentCount": 6,
  "charCount": 343,
  "wordCount": 59,
  "text": "This is a short sample recording for the transcription actor. The Riverside Community Garden will hold its spring planting day on Saturday, 12th of April, from 9 in the morning until noon. Please bring gloves, a trowel, and a bottle of water. ...",
  "segments": [
    {"id": 0, "start": 0, "end": 4, "text": "This is a short sample recording for the transcription actor."},
    {"id": 1, "start": 4, "end": 8.4, "text": "The Riverside Community Garden will hold its spring planting day on Saturday,"}
  ],
  "subtitles": {
    "srt": "https://api.apify.com/v2/key-value-stores/<store>/records/file-1.srt?signature=...",
    "vtt": "https://api.apify.com/v2/key-value-stores/<store>/records/file-1.vtt?signature=..."
  }
}

A failed file has "status": "error" and "error": {"code": "no-audio-track", "message": "..."}. A truncated file has "truncated": true and a truncatedReason: maxMinutesPerFile, the run's maximum charge or

the run's time limit

Pricing

Pay per event, no start fee, and nothing for files that fail:

Each file is billed by its started minutes of audio, at least 1 minute per file: a 30-second voice note costs $0.01, a 4 minute 10 second clip costs $0.05, and a one-hour meeting costs $0.60. Only the audio actually transcribed is billed. When you set a maximum charge for a run, the Actor stops before a file it cannot pay for, and transcribes only the minutes the remainder covers (the item is marked truncated).

Limits

• Speed: about 2 minutes of audio per minute of run time at the default 1 GB of memory, on clear speech. Noisy audio is slower. If the run's timeout comes first, the Actor stops a minute early, saves what it transcribed (marked truncated) and bills only that; give long files a longer run timeout.
• There is no speaker labelling (diarization): the text does not say who spoke.
• base trades accuracy for speed and price. Names, rare words and numbers can be misheard; numbers are usually written as digits.

Licences

The speech model and libraries are open source: Whisper's code and weights and faster-whisper are MIT, CTranslate2 is MIT (it includes Intel oneMKL under the Intel Simplified Software License), PyAV is BSD-3-Clause, and this software uses libraries from the FFmpeg project under the LGPLv2.1, built LGPL-only and audio-only from FFmpeg's release source. The full list, with licence texts and sources, is in the Actor's THIRD_PARTY_NOTICES file.