Subtitle Generator - SRT / VTT from audio & video (Whisper)

Turn any audio or video file into clean, readable subtitles with OpenAI's Whisper running on CPU inside Apify. No API key, no sign-up with a third party, no file ever leaves the Apify platform.

What makes this one different from a plain transcriber:

• Broadcast-style cues. Subtitles are rebuilt from word timestamps into cues that respect a maximum line length, a maximum number of lines and a maximum duration, and break at sentence ends and pauses, instead of dumping Whisper's raw segments.
• CJK-aware. Japanese, Chinese and Korean have no spaces; wide characters are counted properly, lines break after punctuation, and cues follow Whisper's phrase boundaries. maxCharsPerLine: 32 gives tidy 16-character Japanese lines for vertical video.
• Bilingual subtitles. One click adds an English translation under every original line - for language learning content, travel vlogs and interviews.
• Word-level timestamps on request, for karaoke-style or word-highlight captions.
• Whisper large-v3-turbo on the Accurate tier: near large-v3 accuracy at a fraction of its cost. Japanese narration comes out with correct proper nouns and punctuation (a punctuation prompt is applied automatically for Japanese).
• Three quality levels (small greedy / small / large-v3-turbo) so you can trade speed for accuracy per job.

Input

Field	Type	Default	Description
mediaUrls	array	-	Direct links to mp3, wav, m4a, ogg, flac, mp4, mov, webm, mkv... (anything FFmpeg decodes). YouTube / TikTok page links are not supported; pass a direct file URL.
language	string	auto	ISO 639-1 code (en, ja, es...). Empty = auto-detect.
quality	enum	balanced	fast (small, greedy), balanced (small, default), accurate (large-v3-turbo, ~3x slower, best for Japanese/Chinese/accents).
translateToEnglish	boolean	false	Adds English transcript + English SRT/VTT (second pass).
bilingualSubtitles	boolean	false	Extra SRT with original line + English line in every cue.
maxCharsPerLine	integer	42	Wide (CJK) characters count double. 32 for mobile/vertical.
maxLinesPerCue	integer	2
maxCueDurationSec	integer	6
wordTimestamps	boolean	false	Adds a words array per file.
vocabularyHint	string	-	Names and jargon, comma separated.
vadFilter	boolean	true	Skip silence; avoids hallucinated text.
saveFilesToKeyValueStore	boolean	true	Store .srt, .vtt, .txt files and return their URLs.
maxDurationMinutes / maxFileSizeMb / downloadTimeoutSec	integer	180 / 500 / 120	Safety limits; rejected files are not charged.

Example

{
  "mediaUrls": ["https://example.com/interview.mp4"],
  "language": "ja",
  "quality": "accurate",
  "bilingualSubtitles": true,
  "maxCharsPerLine": 32
}

Output

One dataset record per file:

Field	Description
url, fileName, ok, error	Source and status. Failed files get ok: false and an error message.
language, languageProbability	Detected language.
durationSeconds, tookMs	Media length and processing time.
text	Full transcript.
segments	Whisper's raw segments (start, end, text).
cues	The subtitle cues after re-flow (start, end, text).
srt, vtt	Subtitle files as text.
srtUrl, vttUrl, txtUrl	Download links (key-value store).
words	Word-level timestamps (when wordTimestamps is on).
translationText, translationSrt, translationVtt, translationSrtUrl	English translation (when requested).
bilingualSrt, bilingualSrtUrl	Original + English cues (when requested).

Pricing

You pay per started minute of audio that was transcribed successfully. accurate quality costs more per minute, and translation / bilingual output adds a second per-minute charge. Downloads that fail, files above the size or duration limits and undecodable inputs are not charged. Platform usage is included.

Speed and limits

• fast runs at roughly 3x realtime on the default 4 GB memory (a 30-minute podcast in about 10 minutes); balanced about 1.5x; accurate about 1x realtime. Give long files more memory (8 GB = 2 CPUs) to finish sooner at the same price.
• English translation uses Whisper medium (large-v3-turbo cannot translate), so translated runs take about twice as long.
• Timestamps are word-level for Latin scripts and character-level for CJK scripts.
• No speaker diarization. Overlapping speech and heavily compressed phone audio reduce accuracy.
• Links must be public. Signed URLs that expire in seconds may fail to download.

Tips

• For vertical videos (Shorts, Reels, TikTok) use maxCharsPerLine: 32 and maxLinesPerCue: 2.
• Put proper nouns in vocabularyHint; Whisper spells them far better with a hint.
• For Japanese use quality: "accurate", set language: "ja" explicitly and put names in vocabularyHint.
• Chain this Actor with a video downloader Actor if your source is a YouTube or TikTok page.

Support

Open an issue on the Issues tab with the media URL (or its format and length) and the input you used.