Audio & Video Transcriber — Speech to Text, SRT & Timestamps

Turn any audio or video file into accurate, timestamped text. Whisper runs inside the Actor — no API key, no external service, nothing to sign up for.

Built for the recordings that break other transcribers: long ones.

---

Why long recordings break most transcribers

Whisper is excellent on a clean three-minute clip. On a two-hour podcast, lecture or meeting it has three well-documented failure modes, and most transcription tools ship with none of them handled:

What goes wrong	What you get back
Repetition loops	The same clause emitted hundreds of times. A 90-minute recording returns "Thanks for watching. Thanks for watching. Thanks for watching…" for forty minutes.
Silence hallucination	Dead air, music beds and room tone produce confident, fluent sentences that nobody said.
Timestamps swallowing silence	One six-word line stamped as lasting 50 seconds, because a pause got absorbed into it. Your subtitle sits on screen for a minute and your billing counts silence as speech.

This Actor detects and repairs all three — and tells you when it did, in a repairs field. A transcript that quietly threw away a minute of audio is worse than one that says so.

---

What you get

Accurate text, in the shape you need

• Plain text — the whole transcript.
• Paragraphs — split at natural pauses, so an hour-long transcript is actually readable instead of one wall of words.
• SRT and WebVTT subtitles — correctly formatted, with non-overlapping cues that players will not flicker on.
• Timestamped segments — every recognised span with start and end times.

Timestamped chunks for RAG (optional)

Turn on chunkForRag and the transcript comes back as retrieval-ready chunks:

{
  "index": 7,
  "timestamp": "00:12:34 - 00:13:10",
  "start": 754.0,
  "end": 790.0,
  "text": "The second thing we changed was the onboarding flow..."
}

• A chunk never splits a sentence, so boundaries never land mid-thought.
• Chunks prefer to break at a real pause, graded by how long the silence was — a 20-second gap is a topic change, a 1-second one is a breath.
• Every chunk carries its timestamp, so a search hit can be cited or played back at the exact second it was said.

That last point is the difference between a transcript you can search and one you can use. Most tools hand you text with no way back to the audio.

Automatic language detection

90+ languages, detected automatically, or pin it explicitly with language for a faster and more reliable run.

Built for batches

One row per file. A 404, an unreadable file or a recording with no speech returns a row with status: "error" (or a clear warning) and never aborts the run.

---

How to use it

1. Paste one or more direct media URLs into Media URLs. Anything you can link to works — a file on your own server, an S3 or Drive direct link, a podcast episode enclosure.
2. Pick a model. base is the default and handles clean speech well; choose small for heavy accents, background noise or overlapping speakers.
3. Choose your output formats — plain text, paragraphs, SRT, VTT, or several at once. Turn on RAG chunks if the transcript is going into a vector database.
4. Start the run. Results appear as one dataset row per file, with the transcript, timestamps and subtitles together.

Every run is a normal Apify run, so you can schedule it, trigger it from the API, or wire it into Make, Zapier, LangChain or LlamaIndex.

Pricing — you are charged for speech, not for silence

Billed per minute of detected speech, rounded up per file.

• A two-hour recording containing five minutes of talking bills five minutes, not 120.
• Silence, music-only and room-tone files bill nothing.
• Failed downloads and unreadable files bill nothing.
• A run that transcribes nothing costs nothing.

Most transcribers bill the full length of the file regardless. Paying for dead air on a recording you did not edit is not a service you should have to buy.

---

Input

Field	Type	Default	Description
mediaUrls	array	—	Required. Direct links to audio/video files or podcast enclosures.
model	string	base	tiny, base or small. Larger is more accurate on hard audio.
language	string	auto	ISO code (en, nl, de…). Empty = detect.
outputFormats	array	["text"]	Any of text, paragraphs, srt, vtt.
includeSegments	boolean	true	Timestamped spans.
chunkForRag	boolean	false	Also return timestamped retrieval chunks.
chunkSize	integer	1200	Soft character ceiling per chunk.
chunkOverlap	integer	120	Characters repeated between chunks.
maxMinutesPerFile	integer	0	Cost ceiling per file. 0 = no limit.

Example

{
  "mediaUrls": [
    "https://example.com/episode-42.mp3",
    "https://example.com/all-hands.m4a"
  ],
  "model": "base",
  "outputFormats": ["text", "srt"],
  "chunkForRag": true
}

Output

{
  "url": "https://example.com/episode-42.mp3",
  "filename": "episode-42.mp3",
  "status": "ok",
  "language": "en",
  "durationSeconds": 3612.4,
  "speechSeconds": 3380.1,
  "billedMinutes": 57,
  "model": "base",
  "text": "Welcome back to the show...",
  "srt": "1\n00:00:01,140 --> 00:00:04,140\nWelcome back to the show.\n...",
  "segments": [ { "index": 0, "start": 1.14, "end": 4.14, "text": "Welcome back to the show." } ],
  "chunks": [ { "index": 0, "timestamp": "00:00:01 - 00:01:12", "text": "..." } ],
  "repairs": ["trimmed silence inside 812.0-869.0s (57.0s of timeline for 31 characters of speech)"]
}

Failed files:

{
  "url": "https://example.com/missing.mp3",
  "filename": "missing.mp3",
  "status": "error",
  "error": "http 404"
}

---

Supported formats

mp3, m4a, wav, flac, ogg, opus, aac, mp4, webm, mkv, mov and other common containers — the audio track is extracted automatically from video files.

Limitations — stated up front

• No platform downloads. This Actor does not fetch from YouTube, TikTok, Instagram, X or similar sites. It transcribes media you can link to directly. This is deliberate: platform scraping breaks constantly and violates those sites' terms.
• No speaker diarization in this version. Transcripts are not labelled "Speaker 1 / Speaker 2".
• Accuracy depends on the audio. Heavy accents, overlapping speech, poor microphones and background music all reduce quality — use the small model for difficult recordings.
• Text is transcribed in the language spoken; nothing is translated.

Common uses

Podcast and interview transcripts; webinar transcription and lecture transcription; meeting notes; subtitle generation (SRT/VTT); making an audio or video archive searchable; accessibility captions.

For AI pipelines: speech to text RAG ingestion, where every chunk keeps the speech timestamps it came from — so an answer can cite the moment something was said, not just the file it was said in. Building a RAG knowledge base from recordings, or feeding spoken content to an LLM with timestamps intact.

Any direct media URL works, so you can transcribe MP3, WAV, FLAC, M4A, MP4, MOV and the other containers listed under Supported formats.

Integrations

Standard Apify output: pull results via the API, export to JSON/CSV/Excel, schedule runs, or connect to Make, Zapier, LangChain, LlamaIndex and other Apify integrations.

FAQ

Is this legal? Yes. You supply media you already have the right to use, and the Actor transcribes it. No third-party website is accessed, no terms of service are involved, and no proxies are used. That is a deliberate design choice, not an oversight — see the next answer.

Can it transcribe a YouTube, TikTok or Instagram link? No, and it will not be added. Downloading from those platforms violates their terms, needs paid proxies to work at all, and breaks whenever the platform changes. If you need media from a social platform, fetch it with a tool built for that and pass the resulting file URL here.

How much will a run cost? You are charged per minute of detected speech, not per minute of file length, so silence, music beds and dead air are free. A 45-minute interview with 40 minutes of talking costs $1.60. A 3-minute voice note costs $0.12. Failed files cost nothing at all.

What happens if one file in my batch fails? Only that file. It appears in the dataset with status: "error" and the reason, the rest of the batch continues, and you are not charged for it.

Why is my transcript shorter than the recording? Check the warning field on that row. Either a per-file minute cap was set, or the run's maximum cost would have been exceeded — the Actor stops rather than transcribing audio it cannot bill for, and tells you which limit applied.

Does it label who is speaking? No. Speaker diarization was built and tested against this Actor and reached only 50–77% accuracy on real recordings, so it was not shipped. A transcript where a quarter of the lines carry the wrong speaker is worse than one with no labels — and charging for it would be worse still.

Can I get the data out programmatically? Yes — the standard Apify dataset API, plus JSON, CSV and Excel export. Every output field is described in the dataset schema, so tools and AI agents can read the structure without guessing.

Other Actors by this author

Actor	What it does
PDF Text Extractor — Markdown, Tables & RAG Chunks	The same idea for documents: PDF to clean Markdown with real tables, heading structure and RAG-ready chunks. Pairs naturally with this Actor when a knowledge base mixes recordings and documents.