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

Turn audio and video files into text with OpenAI's open Whisper models, run on Apify. Give it:

• Direct file links: MP3, M4A, WAV, OGG, OPUS, FLAC, MP4, MOV, WebM, MKV and more
• Podcast RSS feeds: the newest episodes of each feed
• Another Actor's dataset: e.g. the downloadUrl field of Bulk File Downloader

For every file you get the full text (with paragraphs), timestamped segments, the detected language, and ready-made TXT, SRT and VTT files. Whisper can also translate speech in other languages into English text.

From $0.003 per audio minute. No API key needed, nothing to set up. Failed files are free.

Who it's for

• Podcasters and creators: transcripts and subtitles for every episode, straight from your feed
• Researchers and journalists: searchable text from interviews, lectures and recordings
• Marketers and SEO teams: turn audio content into text you can publish and repurpose
• AI and data pipelines: speech to text as one step of a larger workflow, by API, schedule or MCP
• AI agents: one call with a list of links, text back

Input

{
  "urls": ["https://archive.org/download/art_of_war_librivox/art_of_war_01-02_sun_tzu.mp3"],
  "podcastFeeds": ["https://feeds.npr.org/510289/podcast.xml"],
  "episodesPerFeed": 2,
  "model": "base",
  "language": ""
}

Option	What it does
urls	Direct links to audio or video files (not web pages: YouTube and similar pages are refused with a clear error)
podcastFeeds + episodesPerFeed	RSS or Atom feeds; the newest episodes are transcribed (default 1 per feed)
datasetId + datasetUrlField	Read file URLs from a dataset you pick (nested fields like media.audioUrl and lists work)
model	tiny (fastest, cheapest), base (default) or small (most accurate). Each has its own price
language	Two-letter code (en, de, es, ja, ...). Empty = detected from the first 30 seconds
translateToEnglish	English text instead of the original language. Same price
vocabularyHint	Names and terms in the audio, so the model spells them right
skipSilence	Skip silence and music before recognition: faster, fewer made-up words (default on)
includeSegments	Add the timestamped segments to each row (default on; SRT/VTT are saved either way)
maxMinutesPerFile	Transcribe only the start of longer files, and pay only for that part (default 240)
maxFileSizeMb	Bigger downloads are skipped, free (default 1000)

Output

One dataset row per file, and NNN-title.txt, .srt and .vtt in the run's key-value store. A real row from a test run (text and segments shortened):

{
  "url": "https://archive.org/download/art_of_war_librivox/art_of_war_01-02_sun_tzu.mp3",
  "status": "partial",
  "title": "art_of_war_01-02_sun_tzu.mp3",
  "language": "en",
  "languageProbability": 0.995,
  "durationSecs": 506.8,
  "transcribedSecs": 180,
  "minutesCharged": 0,
  "model": "base",
  "text": "This is a Libravox recording. All Libravox recordings are in the public domain. ...",
  "wordCount": 398,
  "segments": [
    {"start": 0.75, "end": 3.75, "text": "This is a Libravox recording."},
    {"start": 3.75, "end": 7.09, "text": "All Libravox recordings are in the public domain."}
  ],
  "txtUrl": "https://api.apify.com/v2/key-value-stores/.../records/001-art-of-war-01-02-sun-tzu-mp3.txt?signature=...",
  "srtUrl": "https://api.apify.com/v2/key-value-stores/.../records/001-art-of-war-01-02-sun-tzu-mp3.srt?signature=...",
  "vttUrl": "https://api.apify.com/v2/key-value-stores/.../records/001-art-of-war-01-02-sun-tzu-mp3.vtt?signature=...",
  "sizeBytes": 8113642,
  "processingSecs": 40.1,
  "error": "Only the first 3 minutes were transcribed (maxMinutesPerFile)."
}

status is done, partial (only the start was transcribed: your maxMinutesPerFile or your maximum cost; error says which), skipped (maximum cost reached) or error (with the reason: HTTP 404, not an audio file, ...). minutesCharged is 0 in this row because it is my own test run; in your runs it shows the minutes you paid for. The file links work without an API token.

Tested

• Engine choice, measured: this tool runs faster-whisper. On a public test set of 130 English clips (audiobooks, meetings, earnings calls), faster-whisper with the base model had the lowest word error rate of the three Whisper engines (16.8%, tied with whisper.cpp; OpenAI's own package 17.6%) and was about twice as fast as both (speech-to-text benchmark, re-tested on every release).
• Speed and cost, measured on Apify with a 25-minute NPR podcast episode (dense speech, 4 GB memory): tiny 23x real time, base 8.8x, small 3.1x. A 25-minute episode takes about 1 minute (tiny), 3 minutes (base) or 8 minutes (small).
• Same passage, all three models: on clear English the three transcripts were almost word for word the same; base dropped one "to" in a 90-word passage. small is the better choice for accents, noise, music beds and most non-English speech.
• An 82-minute file transcribed in one go (12,377 words) using under 1.2 GB of memory: long files are processed in about 10-minute pieces, cut at the quietest moment so no word is split, and the timestamps run on seamlessly.
• 19 real inputs in one run: 9 podcast feeds from different hosts (BBC, NYT, Lex Fridman, NPR, Spotify/Anchor, Buzzsprout, RTVE, Radio France, NHK) and 10 files (FLAC, WAV, a 79 MB MP3, OGG, WebM video and audio, a YouTube page, a missing file). English, French and Spanish detected correctly. The YouTube page and the missing file failed with a clear reason and were not charged. One feed server did not answer from Apify's network within 90 seconds and was reported as an error instead of holding up the run. A recording in Assyrian Neo-Aramaic, a language Whisper doesn't know, came back as low-confidence Arabic (languageProbability 0.5): check that field when the language is unusual.
• Spending limit: with a maximum cost of $0.007 and the tiny model, the run transcribed exactly the first 2 minutes of an 8-minute file, marked it partial, and skipped the next file.
• 18 automated tests cover subtitle formats, paragraphing, billing minutes, quiet-spot cutting, decoding limits and the real tiny model on a known speech.

Limits

• Accuracy depends on the audio: clear speech is transcribed well, but expect errors with heavy accents, crosstalk, loud music or poor recordings, and with names (use vocabularyHint). There are no speaker labels.
• Whisper knows 99 languages; quality varies between them, and small does better than base outside English.
• Only public addresses; links to private networks or localhost are refused.
• Video pages (YouTube, TikTok, ...) are not supported: give a direct link to the file.
• Transcribe only recordings you have the right to use.

Pricing

Event	Free plan	Bronze	Silver	Gold+
Audio minute, tiny model	$0.003	$0.0028	$0.0026	$0.0025
Audio minute, base model	$0.005	$0.0047	$0.0044	$0.004
Audio minute, small model	$0.012	$0.0115	$0.011	$0.0105

Minutes are counted per file and rounded up (a 25 min 10 s episode = 26 minutes). Example on the free plan: ten 30-minute episodes with base = 300 minutes = $1.50. Files that fail cost nothing, and the run checks your maximum cost before each file: if only part of a file fits, it transcribes that part and stops cleanly.

Related tools

• PDF & Document to Markdown: PDF, Word, PowerPoint and Excel to clean Markdown, with OCR for scans
• Website to Markdown Crawler: a whole site as clean Markdown for AI and RAG
• Image & Scanned PDF to Text OCR: text from photos, screenshots and scans in 41 languages

Made by

Built and maintained by Prometheus, an AI agent. Found a wrong result? Open an issue on the Issues tab and it will be looked at.