Audio & Video Transcriber (Whisper, your own files)

Turn audio and video files into text, timestamped segments, or SRT/VTT subtitles. It runs open-source Whisper (faster-whisper, MIT) on CPU inside the Actor, so your audio is not sent to any third-party speech service.

• Input: direct links to media files (mp3, wav, m4a, aac, ogg, opus, flac, mp4, mov, mkv, webm and anything else FFmpeg can decode). Only the audio track of a video is used.
• Output: one dataset item per URL: plain text plus, if you choose, timestamped segments, SRT or WebVTT. Every item says which engine and model (with its exact revision) produced it, whether silence filtering was on, and whether the transcript tripped a quality check.
• Silence and music are filtered out before decoding (voice activity detection, on by default), because Whisper invents text such as "Thank you." on non-speech audio. See Measured.
• Quality flags: each transcript is checked for the known failure shapes (repeated loops, one phrase repeated many times, uniform 30 s segments, long stretches with no text, a missing tail). A tripped check sets quality_degraded: true and says what and where in quality_flags.
• Vocabulary prompt: give names and terms and they come out spelled your way (measured below).
• Long files: audio is decoded in windows, so memory stays flat however long the file is (tested to 70 minutes).
• Pricing: pay per audio minute. See Pricing.

What it does NOT do

• It does not download from web pages. The URL must return the media file itself; a URL that returns a web page fails with "URL returned a web page, not an audio or video file", and one that returns anything else that is not audio fails with "the URL did not return an audio or video file". Neither is charged.
• No speaker diarisation and no word-level timestamps in this version.

Input

Field	Default	Notes
mediaUrls	(required)	Direct media-file URLs
model	small	tiny, base or small. small is built in; tiny and base download in the run that uses them. (medium and large-v3-turbo appear in the measurements below but are not offered at the current price.)
language	auto	ISO code such as en. Empty = detect. Auto-detect listens to only about the first 30 s of speech, so a file that opens with music or another language can be detected wrongly. Set it if you know it.
outputFormat	segments	text, segments, srt or vtt. text is always included.
vocabulary_prompt	empty	Names and terms, spelled as you want them. Only the last 224 tokens are used (Whisper's limit): put the important terms at the end. A prompt that does not match the audio can bleed into the transcript.
vadFilter	true	Silero voice-activity detection (faster-whisper's vad_filter).
vadThreshold, vadNegThreshold, vadMinSpeechDurationMs, vadMaxSpeechDurationS, vadMinSilenceDurationMs, vadSpeechPadMs	0.5, auto, 0, none, 2000, 400	faster-whisper's own defaults, each explained in the input form
engine	self_hosted	The only engine: Whisper inside the Actor.
windowSeconds	600	Decoding window, 30 to 3600
beamSize	5	1 = greedy (faster)
carryContext	true	Prompt each window with the previous window's last ~200 characters. Keeps punctuation and quieter speakers across window boundaries (measured on a 2-speaker interview: 3.1% vs 5.4% WER off)
maxDownloadMb / downloadTimeoutSecs / cpuThreads	2000 / 300 / 0 (auto)

Output

{
  "url": "https://.../episode-12.mp3",
  "status": "ok",
  "engine": "self_hosted",
  "model": "small",
  "modelRevision": "Systran/faster-whisper-small@536b0662742c02347bc0e980a01041f333bce120",
  "vad": {"enabled": true, "threshold": 0.5, "min_speech_duration_ms": 0, "min_silence_duration_ms": 2000,
          "speech_pad_ms": 400, "speechSeconds": 797.43},
  "quality_degraded": false,
  "quality_flags": [],
  "language": "en",
  "durationSeconds": 802.54,
  "transcribedSeconds": 802.54,
  "chargedUnits": 14,
  "text": "One dollar and eighty-seven cents. That was all. ...",
  "segments": [{"start": 0.0, "end": 3.6, "text": "One dollar and eighty-seven cents."}]
}

status: ok; partial (your maximum charge was reached or the run was aborted: the transcript covers what was paid for); failed (404, a web page or non-media file, or a decoding error part-way: then the transcript covers only the audio before the error and the error says so); rejected (not an http(s) URL, or a private address); not_processed (did not fit in your maximum charge, or the run had stopped).

quality_flags entries are {"gate", "detail"}. Gates: loops, repeats, uniform, gaps, coverage (thresholds from an established transcription checklist: 3 identical segments in a row, one text more than 25 times, 10+ segments of 28–31 s, more than 8 s with no text, the last segment ending 30 s or more before the audio does), plus decode_errors and window_failed. quality_degraded is null for items that were never transcribed. gaps also trips on real pauses longer than 8 s (for example 90 s of silence between two talks); the detail says where, so you can judge.

An item larger than Apify's dataset limit (9 MB of JSON, Apify docs) keeps its metadata in the dataset and stores the transcript in the run's key-value store, linked by transcriptKey / transcriptUrl.

Pricing

One price per minute of audio transcribed, shown on the Pricing tab; no start fee. A partial minute at the end of each file is rounded up. Rejected and failed URLs are not charged. If you set a maximum cost per run, a file that does not fit is skipped without charge (not_processed); if the limit is reached mid-file you get the transcript up to the last paid minute (partial). The price is the same for every model, and larger models take longer.

Measured (this Actor's own tests, 2026-09-30)

One shared 16-vCPU KVM host (Haswell, no AVX-512), often busy with other work. Every figure is one run (n=1), and Apify's hardware differs. WER uses Whisper's English normaliser.

Accuracy depends heavily on the audio

Word error rate by condition and model, using the Actor's default settings (600 s windows, beam 5, VAD on, language en, no vocabulary prompt). On this set, small (the default) ranged from 0.3% on clean synthetic speech to 20.7% on a clip full of unusual names. Real recordings (phone calls, crosstalk, heavy accents, poor microphones) were not tested and may be worse.

Condition	tiny	base	small	medium	large-v3-turbo
TTS, clean (6 voices)	1.4%	0.7%	0.3%	0.3%	0.6%
LibriVox story, one reader	5.1%	3.7%	2.0%	1.7%	1.1%
LibriVox sonnet, 17 readers pooled [range]	9.2% [4.0-16.9]	6.7% [3.2-12.1]	5.0% [3.2-8.1]*	8.4% [3.2-32.3]*	3.6% [1.6-7.3]*
TTS + pink noise, 20 dB	1.2%	1.2%	0.4%	0.3%	0.2%
TTS + pink noise, 10 dB	2.2%	1.4%	0.4%	0.3%	0.2%
TTS + pink noise, 5 dB	4.4%	3.5%	0.9%	0.9%	0.6%
TTS + pink noise, 0 dB	18.5%	9.9%	4.3%	2.2%	1.8%
TTS + music, 10 dB	1.4%	0.9%	16.5%*	0.2%	0.2%
TTS + music, 0 dB	9.7%*	7.9%*	7.2%	10.7%*	10.2%*
LibriVox story + pink noise, 10 dB	7.9%	5.4%	3.0%	1.7%	1.5%
LibriVox story + music, 10 dB	6.9%	4.9%	2.7%	2.2%	1.5%
TTS + competing talker, 10 dB	11.1%	4.6%	1.1%	0.6%	4.5%
TTS through a phone channel	10.9%*	8.8%*	7.8%	7.0%*	7.2%*
Unusual proper nouns, no prompt	29.6%	28.1%	20.7%	18.5%	20.7%

WER in percent, lower is better; VAD on, n=1 run per cell. * = at least one quality gate tripped on that cell (see quality_flags in results/accuracy.jsonl).

How the set was built (only audio we have rights to):

• LibriVox readings are public domain: The Gift of the Magi (one reader, 13.4 min) and Shakespeare's Sonnet 130 read by 17 different volunteer readers (different voices, accents and microphones, 56–91 s each). The sonnet reference is the Project Gutenberg text; its "damask'd" normalises to "damask would", so each reader carries up to two unavoidable errors (the sonnet figures are upper bounds). Readers' spoken introductions are not scored.
• TTS: our own 5.5-minute meeting script read by six Kokoro-82M voices (two British, four American). Clean synthetic speech is the easy case.
• Noise and music were synthesised (pink noise; a chord-and-melody loop) and mixed at the stated signal-to-noise ratio (speech level over its non-silent frames vs. the added sound). Talker: the LibriVox reading mixed 10 dB under the TTS speech. Phone: a synthetic phone channel (300–3400 Hz, 8 kHz, 8-bit mu-law).
• tech_clean: a 1-minute Kokoro clip with 15 invented or rare proper nouns (24 occurrences), scored without a vocabulary prompt. With the prompt, see below.
• A notable failure: small at music 10 dB dropped a whole ~120-word paragraph (and inserted a near-duplicate sentence). This happened with VAD on and with VAD off. The gaps check flagged it.

Silence and music (VAD)

Synthetic test file: 120 s of music, 45 s of speech, 90 s of digital silence, 45 s of speech. Segments whose midpoint falls in the music or silence are invented text:

model	VAD off	VAD on (default)
large-v3-turbo	22 invented segments ("Thank you." ×6 in 30 s blocks, then invented sentences such as "What do we do with the new site?"), loops flag	0
small	3 × "you" in the silence	0
tiny	9 ("Thanks for watching." ×2, then repeated real sentences placed in the silence)	0

Pure silence (60 s): "you" ×2 with VAD off for every model, nothing with VAD on. Pure music (120 s): large-v3-turbo wrote "Thank you." ×4 with VAD off, nothing with VAD on. The cost: on the clean LibriVox reading, small scored 1.77% WER with VAD off and 1.96–2.0% with it on.

Vocabulary prompt

One-minute clip with 15 unusual names and terms (24 occurrences), e.g. "Oleksandra Wyrzykowski", "Zephyrine", "Hraunfoss". Exact spelling counts.

model	no prompt	with prompt (80 tokens)	terms first, then filler (464 tokens, cut to last 224)	filler first, terms last (465 tokens, cut to last 224)	prompt words in unrelated audio
tiny	1/15 terms, 2/28 occurrences, WER 29.6%	9/15 terms, 16/28 occurrences, WER 17.8%	1/15 terms, 2/28 occurrences, WER 31.1%	10/15 terms, 18/28 occurrences, WER 13.3%	0
base	0/15 terms, 0/28 occurrences, WER 28.1%	13/15 terms, 25/28 occurrences, WER 7.4%	0/15 terms, 0/28 occurrences, WER 32.6%	12/15 terms, 23/28 occurrences, WER 8.9%	0
small	2/15 terms, 4/28 occurrences, WER 20.7%	14/15 terms, 26/28 occurrences, WER 1.5%	3/15 terms, 6/28 occurrences, WER 23.0%	14/15 terms, 27/28 occurrences, WER 3.7%	0
medium	3/15 terms, 6/28 occurrences, WER 18.5%	15/15 terms, 28/28 occurrences, WER 0.0%	3/15 terms, 6/28 occurrences, WER 18.5%	15/15 terms, 27/28 occurrences, WER 0.7%	0
large-v3-turbo	4/15 terms, 8/28 occurrences, WER 23.0%	14/15 terms, 26/28 occurrences, WER 2.2%	5/15 terms, 11/28 occurrences, WER 12.6%	15/15 terms, 28/28 occurrences, WER 5.9%	0

Terms first then filler: the prompt is over 224 tokens, so the terms are cut off and it behaves like no prompt. Filler first, terms last: the terms survive the cut. The 5th column counts prompt words that appeared in a different recording containing none of them.

The same prompt, applied to a different recording with none of those words, put 0 of them into the transcript for all five models (the bleed test). That is one test, not proof that bleed cannot happen.

Memory

Peak memory depends on the window length and the model, not on file length (70-minute file, small, 600 s windows: 1.6 GB for the whole Actor process). large-v3-turbo peaked at 2.1 GB on the Actor's own path, so it needs the 4 GB setting; it does not fit 2 GB.

Rights and responsible use

Only submit media that you own or are licensed to process, with any consent the recording requires.

About this Actor

Built, tested and maintained with an AI coding agent working for its publisher. Dependencies: faster-whisper 1.2.1 (MIT), CTranslate2, PyAV 15.1.0 (bundled FFmpeg), Silero VAD (via faster-whisper), OpenAI Whisper weights converted to CTranslate2 (MIT; revisions pinned and checked at build), Apify SDK for Python 4.0.2.