Video to Captions (SRT, VTT, TXT) - Whisper AI

Video to SRT, VTT and plain-text transcripts in one API call: send a video or audio URL and ENT-E's Whisper transcription API returns ready-to-upload caption files with word-level timestamps in 99 languages. Nothing to install and no API key from a third party.

Video to SRT and VTT subtitles with a Whisper transcription API

Each file gives you four downloads:

• an .srt file for YouTube, Premiere, DaVinci Resolve or any editor
• a .vtt file for HTML5 players
• a clean .txt transcript that's ready for an LLM
• a .json file with every segment and word, with start/end times and confidence

The transcript text is also returned in the dataset row, so an agent can read it without downloading anything.

Captions are line-broken for reading out of the box: 42 characters per line, 2 lines per cue and at most 6 s per cue. Every setting is adjustable, down to 1–4 words per cue for TikTok-style pop-on captions.

What you get

• Auto captions API for any media URL. MP4, MOV, MKV, WebM, MP3, WAV, M4A and anything else FFmpeg reads, up to 2 GB.
• Whisper models run with faster-whisper. Pick by speed or accuracy:
  • tiny, base and small (the default), at the standard rate.
  • English-only variants tiny.en, base.en and small.en.
  • large-v3-turbo for accented, noisy, technical or non-English speech, at the HQ rate.
• Auto language detection (or set language to an ISO 639-1 code such as en, es, de), and translate to English with "task": "translate".
• Trim before you pay. startSeconds and durationSeconds transcribe only the part you need. Timestamps stay relative to the original file, so the captions still line up.
• Voice-activity filter (vadFilter, on by default). It skips silence and music, so Whisper doesn't invent text there.
• Vocabulary hints. initialPrompt helps the model spell names of products, brands and jargon.
• Batch with files, with per-file overrides.
• Failed files are free.

Use cases

Video to SRT for YouTube uploads

{
  "url": "https://example.com/videos/launch-keynote.mp4",
  "language": "en"
}

Upload srtUrl as the subtitle track. Setting language explicitly is faster than auto-detect and avoids misdetection on short clips.

VTT subtitles for a web player

{
  "url": "https://example.com/videos/onboarding-01.mp4",
  "maxCharsPerLine": 37,
  "maxLinesPerCue": 2
}

Use vttUrl in an HTML5 <track> element.

Word-level timestamps for editing, search and short-form captions

The .json output lists every word with start, end and confidence. For bold one-to-three-word captions:

{
  "url": "https://example.com/clips/interview-cut.mp4",
  "wordsPerCue": 3,
  "maxLinesPerCue": 1,
  "maxCueSeconds": 2
}

Transcribe video API for LLM pipelines (RAG, summaries, show notes)

{
  "url": "https://example.com/podcast/ep-88.mp3",
  "model": "small",
  "txtTimestamps": false
}

Read text from the row, or fetch txtUrl. The text is in clean paragraphs that break after pauses longer than 2 s. Set "txtTimestamps": true for [HH:MM:SS] lines.

Whisper transcription API for hard audio, with English translation

{
  "url": "https://example.com/videos/field-report-es.mp4",
  "model": "small",
  "task": "translate",
  "initialPrompt": "ENT-E, Kubernetes, PyTorch, Acme Robotics"
}

translate outputs English captions for non-English speech. Use it with tiny, base or small; the English-only models and large-v3-turbo aren't trained for it. For the best same-language accuracy on noisy, accented or technical audio, use "model": "large-v3-turbo".

Batch captions for a course or channel back catalog

{
  "model": "small",
  "language": "en",
  "files": [
    "https://example.com/course/lesson-01.mp4",
    "https://example.com/course/lesson-02.mp4",
    { "url": "https://example.com/course/lesson-03.mp4", "startSeconds": 30, "durationSeconds": 600 }
  ]
}

Pricing (pay per event)

Event	Price
Transcript created (per successful file)	$0.005
Audio minute: tiny, base, small and .en models (per started minute transcribed)	$0.02
Audio minute HQ: large-v3-turbo (per started minute, instead of the standard rate)	$0.06
Actor start (per run, one event per GB of memory, minimum one; default 4 GB = 4 events)	$0.00005 per event

There are no platform usage fees on top. Failed files are free.

Worked example: captions for a 9 min 40 s video

• The video is 580 s, which is 10 started minutes.
• With small: $0.005 + 10 × $0.02 = $0.205.
• With large-v3-turbo: $0.005 + 10 × $0.06 = $0.605.

A one-hour podcast: $0.005 + 60 × $0.02 = $1.205 with small, or $3.605 with large-v3-turbo.

Trim to save: captioning only a 2-minute highlight ("startSeconds": 1200, "durationSeconds": 120) costs $0.005 + 2 × $0.02 = $0.045.

Billing counts started minutes of the audio that's actually transcribed: the trimmed range if you set one, otherwise the whole file. It rounds up, ignoring up to 0.5 s of padding, with a minimum of 1. Files with no detected speech are free. Set a spending limit on the run to cap the total.

Input reference

Field	Type	Default	Notes
url	string	(none)	Direct link to a video or audio file, up to 2 GB
files	array	(none)	Batch. Items are { "url": ..., <any setting below> } or plain URL strings.
model	string	small	tiny, base, small, large-v3-turbo (HQ rate), tiny.en, base.en, small.en
language	string	auto	ISO 639-1 code or auto (detected from the first 30 s)
task	string	transcribe	transcribe or translate (to English)
startSeconds	number	(none)	Start of the range to transcribe
durationSeconds	number	(none)	Length of the range to transcribe (min 1)
maxCharsPerLine	integer	42	10–120
maxLinesPerCue	integer	2	1–4
maxCueSeconds	number	6	1–30
wordsPerCue	integer	(none)	1–20, a fixed number of words per caption
vadFilter	boolean	true	Skip non-speech
beamSize	integer	5	1–10. Higher is slightly more accurate but slower.
initialPrompt	string	(none)	Vocabulary and spelling hints
txtTimestamps	boolean	false	Adds [HH:MM:SS] lines to the .txt
maxDurationSeconds	integer	10800	Longer audio is rejected (free). Trim with durationSeconds.

You must provide either url or files.

Output reference

Each file produces one dataset row (URLs and text shortened):

{
  "sourceUrl": "https://example.com/videos/launch-keynote.mp4",
  "model": "small",
  "task": "transcribe",
  "srtUrl": "https://api.apify.com/v2/key-value-stores/<storeId>/records/d6a78644e2db.srt?signature=<sig>",
  "vttUrl": "https://api.apify.com/v2/key-value-stores/<storeId>/records/d6a78644e2db.vtt?signature=<sig>",
  "txtUrl": "https://api.apify.com/v2/key-value-stores/<storeId>/records/d6a78644e2db.txt?signature=<sig>",
  "jsonUrl": "https://api.apify.com/v2/key-value-stores/<storeId>/records/d6a78644e2db.json?signature=<sig>",
  "status": "ok",
  "language": "en",
  "languageProbability": 0.996,
  "durationSeconds": 580.17,
  "segmentCount": 162,
  "cueCount": 210,
  "wordCount": 1812,
  "text": "Welcome to the launch. Today we are going to walk through…",
  "input": { "format": "mov,mp4,m4a,3gp,3g2,mj2", "duration": 580.125, "sizeBytes": 57432329, "video": { "codec": "h264", "width": 854, "height": 480 }, "audio": { "codec": "aac", "sampleRate": 44100, "channels": 2 } },
  "billedMinutes": 10
}

text is capped at 200,000 characters in the row; the full transcript is always in txtUrl. If no speech is found (silence or music only), the row has status: "no-speech", billedMinutes: 0 and a warning, and nothing is charged. A failed file returns sourceUrl, model, task, status: "error" and error. The Output tab also has a "text" view with only sourceUrl, language and text.

Use it from AI agents and automations

MCP (Claude, Cursor, other MCP clients). Add https://mcp.apify.com?tools=ent-e/video-to-captions-whisper as a remote MCP server and sign in with OAuth. The agent calls the Actor with a URL, then reads text from the dataset, or passes srtUrl to the next step. With small at the default memory, a 9.7-minute video took about 6 minutes to transcribe in testing, so the agent should wait for the run to finish before reading results.

n8n. Use the Apify node (@apify/n8n-nodes-apify):

• Run an Actor and get dataset for short clips.
• For longer files, use Run an Actor followed by Get Dataset Items.

Make. Use the Apify app:

• Run an Actor with Run synchronously waits up to 120 s, which suits short clips.
• For longer videos, use the Watch Actor Runs trigger, then Get Dataset Items.

Zapier. Use the Apify app:

• Run Actor and Fetch Results for runs under 5 minutes.
• Otherwise, use Run Actor, then the Finished Actor Run trigger and Fetch Dataset Items.

API.

curl -X POST "https://api.apify.com/v2/acts/ent-e~video-to-captions-whisper/runs?token=$APIFY_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"url":"https://example.com/videos/launch-keynote.mp4","language":"en"}'
# then: GET https://api.apify.com/v2/datasets/<defaultDatasetId>/items?token=$APIFY_TOKEN

Short clips can use /run-sync-get-dataset-items, which waits up to 5 minutes.

FAQ

Can I paste a YouTube or TikTok page link? No. The Actor needs a direct link to the media file itself (for example a .mp4 or .mp3 in cloud storage, a CDN or a file host). To get audio out of something else first, use Media Converter (FFmpeg) on the direct file.

Which model should I pick?

• Start with small.
• Use tiny or base for fast rough drafts.
• Use large-v3-turbo for accents, noise, crosstalk, technical terms or non-English speech. It costs three times as much per minute and gives the most accurate same-language captions.

Does it label speakers? No, there's no speaker diarization. You get timed text, not "Speaker 1 / Speaker 2."

Are the timestamps accurate enough for word-by-word captions? Every word in the .json output has its own start and end time and a confidence value. Use wordsPerCue (1–4) for pop-on captions, or feed the JSON to your own renderer. Or let Video to Shorts burn them in for you.

How do I cut cost on long videos?

• Transcribe only the part you need with startSeconds + durationSeconds. You're billed for that range only.
• Stay on small unless you need large-v3-turbo.

How many languages? Whisper covers 99 languages, with auto-detect or a fixed language code. Accuracy varies by language and audio quality, so use large-v3-turbo for harder languages.

Am I charged if a file fails? No. Download errors, unreadable files and over-length files return a free error row.

ENT-E media toolkit

Every tool takes a public file URL, needs nothing installed, charges per file, doesn't charge for failures, and works from MCP, n8n, Make, Zapier and the API.

• Media Converter for AI Agents (FFmpeg): audio/video to MP3, Whisper-ready 16 kHz WAV, MP4, WebM, GIF or thumbnails, with trim and batch.
• Image Converter & Compressor: convert to WebP or AVIF, compress to a target file size, resize, strip EXIF/GPS.
• Podcast Loudness Normalizer: normalize to -16, -14, -23 or -24 LUFS with a true-peak ceiling, plus a before/after report.
• Video to Captions: SRT, VTT, TXT and word-level JSON from any video or audio URL, with Whisper.
• Video to Shorts: vertical 9:16 clips with burned-in animated captions for TikTok, Reels and Shorts.