Audio & Video Transcriber — speech to text, SRT/VTT subtitles and podcast feeds

Give it links to audio or video files, or podcast feeds, and get for each one: the full transcript, timestamped segments, ready-to-save SRT and WebVTT subtitles, the detected language, and optionally word-level timestamps or a translation to English. 99 languages, detected automatically.

Whisper runs inside the Actor. There is no API key to bring, no third-party transcription service your audio is sent to, and no model download at run time: the models are in the image.

Pay per audio minute, and only for transcripts. US$0.008 (fast), US$0.02 (balanced, the default) or US$0.04 (best) per started minute. A link that does not answer, a web page, a picture, a silent file: each leaves one row that says what happened and costs nothing.

What makes it different

• Podcast feeds, not just files. Paste an RSS feed or an Apple Podcasts show link and get the newest episodes transcribed, each row carrying the podcast, the episode title, its date, id and page.
• Monitor mode. onlyNewEpisodes remembers what it delivered to you. Put the Actor on a daily schedule and every new episode is transcribed once and charged once; a run with nothing new ends in 3 seconds and costs nothing.
• Your spending limit is respected to the minute. If the limit pays for 3 more minutes and the next file lasts 7, the first 3 minutes are transcribed, the row says truncated, and the run stops. Measured, not promised: a run capped at US$0.03 on a 7-minute file charged exactly 3 fast minutes.
• Three qualities, three honest prices. Whisper base, small and large-v3-turbo, with the speed of each measured on Apify's own machines (below), so you can choose between a cent and an accent.
• Every failure is a row with a reason, never a crashed run: unreachable (with the HTTP status), not_media ("the link answers a web page, not an audio or video file"), too_large, no_speech, no_episodes.
• Survives platform restarts. If Apify moves a long run to another server, the files already delivered are skipped: nothing is transcribed or charged twice.

Use cases

• Podcasts: transcripts for show notes, SEO pages and search; a scheduled monitor that transcribes every new episode.
• Subtitles: SRT/VTT for videos you host, in the language spoken or translated to English.
• Meetings, interviews, lectures, sales calls recorded to a file: searchable text with timestamps.
• AI pipelines: feed transcripts to an LLM, a RAG index or an agent. One JSON row per file, stable schema.
• After a scraper: another Actor collects media links, this one turns them into text.

Input

field	what it does
mediaUrls	Direct links to files: mp3, m4a, wav, flac, ogg, opus, aac, mp4, mov, webm, mkv… anything FFmpeg decodes. Up to 1,000 per run, 2 GB per file.
podcastFeeds	RSS/Atom feeds, or Apple Podcasts show links (https://podcasts.apple.com/…/id123…), resolved to the show's public feed.
maxEpisodesPerFeed	How many of the newest episodes of each feed (default 1, up to 500).
onlyNewEpisodes	Monitor: skip the episodes this Actor already delivered to you.
quality	fast, balanced (default) or best. See the table below.
language	auto (default) or a code: en, es, de, fr, pt, it, nl, ja, zh, ar, hi, ru… Give it when you know it: a file that opens with music can fool detection.
translateToEnglish	The transcript comes out in English, in the same pass. Not with best.
wordTimestamps	Start, end and confidence of every word. Takes longer.
includeSubtitles	SRT and VTT inside each row (default on).
maxMinutesPerFile	Read only the first minutes of each file; 0 = whole file.

{
    "podcastFeeds": ["https://podcasts.apple.com/us/podcast/the-daily/id1200361736"],
    "maxEpisodesPerFeed": 3,
    "onlyNewEpisodes": true,
    "quality": "balanced"
}

Quality, speed and price

Speeds measured on Apify on 2026-09-20, per 4 GB of run memory (Apify gives one CPU core per 4 GB, and speed scales with it: at 8 GB everything below is twice as fast, at the same price per audio minute).

quality	model	speed at 4 GB	a 60-minute podcast takes	price per audio minute
fast	Whisper base	7.6× real time	~8 min	US$0.008
balanced	Whisper small	2.4× real time	~25 min	US$0.02
best	Whisper large-v3-turbo	1.15× real time	~52 min	US$0.04

fast is fine for clear speech in major languages. balanced is the sensible default. best is for accents, noise, crosstalk and proper names. For long files raise the run's memory to 8 or 16 GB.

Output

One row per file. This is the real row of the file the input is prefilled with (President Johnson, 1968, public domain), with wordTimestamps on and the segments shortened here:

{
    "status": "ok",
    "reason": null,
    "inputUrl": "https://upload.wikimedia.org/wikipedia/commons/6/6d/Address_to_the_Nation_Excerpt.ogg",
    "mediaUrl": "https://upload.wikimedia.org/wikipedia/commons/6/6d/Address_to_the_Nation_Excerpt.ogg",
    "episode": null,
    "contentType": "application/ogg",
    "fileSizeBytes": 301405,
    "quality": "balanced",
    "language": "en",
    "languageProbability": 0.9912,
    "translatedToEnglish": false,
    "durationSeconds": 37.46,
    "transcribedSeconds": 37.46,
    "truncated": false,
    "billedMinutes": 1,
    "text": "I do not believe that I should devote an hour or a day of my time to any personal partisan causes or to any duties other than the awesome duties of this office, the presidency of your country. Accordingly, I shall not seek, and I will not accept, the nomination of my party for another term as your president.",
    "wordCount": 59,
    "segments": [
        {
            "start": 0.02,
            "end": 8.94,
            "text": "I do not believe that I should devote an hour or a day of my time to any personal partisan",
            "words": [
                { "start": 0.02, "end": 0.54, "word": "I", "probability": 0.758 },
                { "start": 0.54, "end": 0.7, "word": "do", "probability": 0.976 }
            ]
        }
    ],
    "srt": "1\n00:00:00,020 --> 00:00:08,940\nI do not believe that I should devote an hour or a day of my time to any personal partisan\n\n2\n…",
    "vtt": "WEBVTT\n\n00:00:00.020 --> 00:00:08.940\nI do not believe that I should devote an hour or a day of my time to any personal partisan\n\n…",
    "processingSeconds": 19.19,
    "transcribedAt": "2026-09-20T04:50:56.495Z"
}

A podcast episode carries episode: { feedUrl, podcast, title, guid, publishedAt, pageUrl }.

field	meaning
status	ok, or why there is no transcript: no_speech, unreachable, not_media, too_large, no_episodes, failed. Only ok is charged.
reason	In plain words: the HTTP status, what the link turned out to be, or how much of the file was read.
text	The transcript as prose; a blank line where the speaker paused two seconds or more.
segments	start, end, text, and words when asked for. Seconds from the start of the file.
srt, vtt	The subtitles, ready to save as files.
language, languageProbability	What was spoken and how sure the engine is.
durationSeconds, transcribedSeconds, truncated	The file's length and how much of it was read.
billedMinutes	What this row cost, in audio minutes of its quality.
wordCount	null for languages written without spaces (Chinese, Japanese, Thai), where a count would mislead.

The run's summary (files per status, minutes billed, episodes the monitor skipped) is in the key-value store as SUMMARY.

Cost

what	audio	price
The prefilled example, balanced	37 s	US$0.02
A 30-minute podcast episode, balanced	30 min	US$0.60
The same, fast	30 min	US$0.24
A 10-hour audiobook, fast	600 min	US$4.80
A daily monitor on a feed with nothing new	—	US$0.00

A file is billed by the started minute, at least one. There is no start fee and no per-file fee. Measured on 2026-09-20: two LibriVox chapters from a feed, 36 minutes of Spanish, balanced, at 4 GB: 15 minutes of run, US$0.72.

Limits

• Direct file links and feeds only. A YouTube, TikTok or Instagram page is not a file; this Actor says so in a free row instead of trying to get around those sites. (For YouTube, captions exist: see our YouTube Transcript Scraper.)
• No speaker labels (diarization).
• Whisper can mishear names and rare words; best mishears fewer. Music and silence are skipped by voice detection.
• 2 GB per file. Files are processed one after another: for many long files, run several runs side by side.
• Feeds that need a login or a private token in a header are not supported; a private feed whose token is part of the URL works.

Is it legal?

The Actor downloads the files you point it at and nothing else. Transcribe recordings you have the right to process: your own, public-domain and openly licensed ones, or ones you have permission for. Transcripts of people's voices can be personal data; if you process them, the GDPR and similar laws apply to you as they would with any other tool. Whisper is open source (MIT licence, by OpenAI); this Actor is not affiliated with OpenAI.