Audio & Video Transcriber: Whisper Speech to Text, SRT, VTT

Speech to text and audio to text for audio, video and podcast episodes: text and subtitles. Paste links to MP3, WAV, M4A, MP4 or other media files, or a podcast RSS feed, and get the full transcript, timed segments and ready-to-use TXT, SRT and VTT files. It runs OpenAI's open-source Whisper model directly on Apify, so your audio is not sent to any third-party AI API (no OpenAI, Groq or Deepgram account needed).

• 99 languages, automatic language detection, or set the language yourself
• Translate speech in any language into English text
• Timed segments, optional word-level timestamps with confidence scores
• Speaker labels (who spoke when): optional speaker diarization labels every segment Speaker 1, Speaker 2, ... and adds a dialogue transcript, speaking time per speaker and speaker names in the SRT and VTT files, all with open models on Apify (no pyannote or Hugging Face token needed)
• Podcast RSS feeds: give the feed link, get the newest episodes transcribed with show and episode title, date and link on each row (schedule it daily to transcribe every new episode)
• Many files in one run; long files up to 4 hours each
• Pay per audio minute; failed files are free
• Empty input = a free short sample (with speakerLabels, a two-voice sample), so you can see the output before paying

Who uses it

• Podcasters and video editors who need subtitles (SRT/VTT) and show notes
• Podcast monitoring and research: transcribe the latest episodes of many shows from their RSS feeds
• Researchers and journalists transcribing interviews (with speaker labels for Q&A)
• Meeting notes and call analysis: who said what, and how long each person spoke
• Teams turning meeting and webinar recordings into searchable text for AI tools and RAG pipelines
• No-code automations (Make, Zapier, n8n) that need speech-to-text from a file link

Input

{
  "audioUrls": [
    "https://example.com/podcast/episode-12.mp3",
    "https://example.com/videos/webinar.mp4"
  ],
  "podcastFeedUrls": ["https://feeds.example.com/my-show.xml"],
  "maxEpisodesPerFeed": 3,
  "language": "",
  "task": "transcribe",
  "model": "base",
  "wordTimestamps": false,
  "speakerLabels": true,
  "numSpeakers": 0
}

• audioUrls: direct links to files (anything FFmpeg can read: mp3, wav, m4a, aac, ogg, opus, flac, webm, mp4, mov, mkv, ...). Files in your own Apify key-value store or cloud storage work through a public or signed link.
• podcastFeedUrls: RSS or Atom feeds of podcasts (the feed link, not the Apple Podcasts or Spotify page). The newest maxEpisodesPerFeed episodes (default 3) of each feed are transcribed. Each row then also has podcastTitle, episodeTitle, episodePublishedAt, episodeUrl, episodeGuid and feedUrl. Owner names and e-mail addresses in feeds are never collected. Feeds and episode files are checked against robots.txt like any other link. Use transcripts of other people's shows in line with their rights (for example for research, search or accessibility), not to republish them.
• language: a code such as en, ja, de, es, or empty to detect it.
• task: transcribe (same language) or translate (to English).
• model: base (default, more accurate) or tiny (faster). Same price.
• speakerLabels (default off): find the speakers and label each segment (segments[].speaker, also on every word with word timestamps). Rows get speakerCount, speakers (speaking seconds, turns, first time heard) and dialogue (Speaker 1: ... paragraphs); the TXT file becomes the dialogue, SRT lines start with [Speaker 1], VTT uses voice tags <v Speaker 1>. numSpeakers: the number of people if you know it (for example 2 for an interview), 0 = automatic. Extra $0.013 per audio minute.
• maxMinutesPerFile (default 120, up to 240) and maxFileSizeMb (default 500) protect your budget.

Pages of YouTube, TikTok, Instagram, Spotify and similar platforms are refused: their terms do not allow downloading, and they are not media files.

Output

One dataset row per file:

{
  "url": "https://example.com/podcast/episode-12.mp3",
  "fileName": "episode-12.mp3",
  "durationSecs": 1834.2,
  "billedMinutes": 31,
  "language": "en",
  "languageProbability": 0.99,
  "text": "Welcome back to the show. Today we talk about ...",
  "wordCount": 5120,
  "segments": [
    {"start": 0.0, "end": 3.2, "text": "Welcome back to the show."},
    {"start": 3.2, "end": 7.9, "text": "Today we talk about ..."}
  ],
  "txtUrl": "https://api.apify.com/v2/key-value-stores/.../records/transcript-00001-episode-12.txt",
  "srtUrl": "https://api.apify.com/v2/key-value-stores/.../records/transcript-00001-episode-12.srt",
  "vttUrl": "https://api.apify.com/v2/key-value-stores/.../records/transcript-00001-episode-12.vtt",
  "truncated": false,
  "processingSecs": 402.5,
  "realTimeFactor": 0.219,
  "error": null
}

With speakerLabels: true each segment also has "speaker": "Speaker 1", and the row adds:

{
  "speakerCount": 2,
  "speakers": [
    {"speaker": "Speaker 1", "speakingSecs": 1010.4, "turns": 58, "firstHeardAt": 0.0},
    {"speaker": "Speaker 2", "speakingSecs": 772.9, "turns": 57, "firstHeardAt": 3.4}
  ],
  "dialogue": "Speaker 1: Welcome back to the show.\n\nSpeaker 2: Thanks for having me. ...",
  "speakerBilledMinutes": 31,
  "speakerError": null
}

Failed files come back as rows with error (for example http_error, not_a_media_file, file_too_large, not_audio, blocked_by_robots_txt; for feeds not_a_feed, feed_http_error, feed_without_media) and are not charged. A SUMMARY record in the key-value store has the totals.

Pricing

$0.006 per started minute of audio, per file ($0.36 per hour), plus Apify's small start fee. A 61-second file is 2 minutes. Files longer than maxMinutesPerFile are cut and billed only up to the cut. Speaker labels (optional): +$0.013 per started minute (so $0.019 per minute with speakers), charged only when turned on and only if the speakers were found (a file whose speaker step fails keeps its transcript and is charged as transcription only). Failed files and the built-in sample are free. You can also set a maximum cost per run in Apify.

Use it from AI agents (MCP)

The Actor is available to AI agents (Claude, ChatGPT, Cursor and others) through Apify's MCP server (mcp.apify.com) and the API. Minimal input (a direct link to an audio or video file):

{"audioUrls": ["https://github.com/ggml-org/whisper.cpp/raw/master/samples/jfk.wav"]}

Output row:

{"fileName": "jfk.wav", "durationSecs": 11, "billedMinutes": 1, "language": "en", "text": "And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country.", "srtUrl": "https://api.apify.com/v2/key-value-stores/.../records/transcript-00001-jfk.srt", "error": null, ...}

How it compares

Popular transcription Actors on Apify Store, with the price each one lists for its main event on 2026-11-05 (check their pages for changes and for other prices):

What you get here for the base price: TXT, SRT and VTT files, timed segments, optional word timestamps, translation to English, podcast RSS feeds, many files per run, failed files free, and audio that never leaves Apify (no third-party AI API). Speaker diarization (who spoke when) is an add-on you pay for only when you use it. Choose a larger-model Actor if you need the highest accuracy on noisy audio: this one runs Whisper tiny/base on CPU.

How it works

1. Each link is checked against the site's robots.txt (on by default) and downloaded once, with a size limit, using an honest User-Agent.
2. The audio is decoded with FFmpeg and transcribed with faster-whisper (CTranslate2, int8, CPU) using the Whisper base or tiny model, with voice activity detection to skip silence.
3. With speaker labels, sherpa-onnx finds the speech turns with the pyannote segmentation-3.0 model, embeds each turn with a 3D-Speaker CAM++ model and groups the turns into speakers; each Whisper segment gets the speaker who talks most during it.
4. The transcript, segments and subtitle files are saved; the downloaded file is deleted.

Speed: the default 4 GB memory gives one CPU core. Each row shows realTimeFactor (processing time divided by audio length); tiny and more memory (8 GB = two cores) are faster for long files.

FAQ

How accurate is it? Whisper base is good for clear speech in common languages; noisy recordings, strong accents, music and overlapping speakers lower accuracy. Setting the language helps.

Does it identify speakers? Yes, with speakerLabels: true: it tells speakers apart and labels them Speaker 1, Speaker 2, ... (in order of first appearance). It does not know people's names; rename them in your own tool. Overlapping speech gets the speaker who talks most in that segment. Setting numSpeakers helps when you know it.

Can I transcribe a YouTube video? No. Use a file you own or may download.

How do I find a podcast's RSS feed? Most podcast websites link it ("RSS"), and podcast hosts (Libsyn, Buzzsprout, Megaphone, Transistor, Podbean, ...) show it on the show page. Directories such as Podcast Index list the feed of almost every show.

Can I get only new episodes every day? Schedule the Actor with maxEpisodesPerFeed: 1 (or the number of episodes a show publishes between runs) and use episodeGuid to skip ones you already have.

Is my audio stored? The downloaded file is deleted after transcription. Transcripts stay in your run's storage under your Apify account.

Limitations

• CPU only; very long files take a while (see Speed above).
• Models: tiny and base. Larger models are too slow on CPU for this price.
• Links must be direct file links that can be downloaded without logging in.

Credits and licences

• OpenAI Whisper models: MIT licence.
• faster-whisper (SYSTRAN) and CTranslate2: MIT licence. FFmpeg (Debian package): LGPL/GPL.
• Speaker labels: sherpa-onnx (k2-fsa, Apache-2.0), pyannote segmentation-3.0 (MIT, converted to ONNX by sherpa-onnx), 3D-Speaker CAM++ speaker embedding model (Alibaba DAMO Academy, Apache-2.0).
• The demo sample is generated with espeak-ng at build time.