From an audio folder to readable subtitles.

A folder of interviews or recorded lessons can become subtitle files without changing your workflow into a video editor. Whisper large-v3 accepts uploaded audio and returns JSON, plain text, SRT, or VTT.

Listen, then compare the transcript.

chapter i the beginning of things things. They were not railway children to begin with. I don't suppose they had ever thought about railways except as a means of getting to Maskelyne and Cook's, the Pantomime, Zoological Gardens, and Madame Tussauds. They were just ordinary suburban children, and they lived with their father and mother in an ordinary, red-brick-fronted villa, with coloured glass in the front door, a tiled passage that was called a hall, a bathroom room with hot and cold water electric bells french windows and a good deal of white paint and

Read by Kara Shallenberg for LibriVox. Text by E. Nesbit. Public domain. Public Domain Mark 1.0. LibriVox states its recordings are public domain in the USA. Recording source. Download the actual JSON output.

The cached sample is real worker output from November 1, 2026, run on a laptop CPU with the real weights. The model missed the spoken introduction before 11.35 seconds and repeated words such as “things things” and “bathroom room.” The output is shown with those mistakes intact.

Ask for SRT at submission.

Follow the upload guide, then set format: "srt" on each item. Set language when you know it, or omit it for detection. SRT is an output format; the cached sample above was run as JSON.

{
  "model": "whisper-large-v3",
  "items": [
    {
      "file": "FILE_ID",
      "params": {
        "language": "en",
        "format": "srt"
      },
      "max_units": 1
    }
  ]
}

The one-minute limit above suits a file no longer than one minute. For a collection, give every uploaded file its own item, measured limit, and params. A batch contains one model, with up to 2,000 items.

Set the hold from the real length.

max_units is an upper bound in audio minutes, not a price quote. The gateway holds the list fare for declared units, or the submission minimum if larger. It releases unused hold at settlement. If the worker measures more than your limit, the item fails without a charge.

Use a local duration probe as an estimate, with some headroom. Billing is based on decoded audio samples rather than the container’s duration field. The smallest allowed bound is 0.5 minutes, and the largest is 240 minutes.

import subprocess

def max_units(path):
    result = subprocess.run(
        ["ffprobe", "-v", "error", "-show_entries", "format=duration",
         "-of", "csv=p=0", str(path)],
        capture_output=True, text=True, check=True,
    )
    minutes = float(result.stdout) / 60
    if minutes > 240:
        raise ValueError("Split this recording before uploading")
    return max(0.5, min(round(minutes * 1.02 + 0.02, 2), 240))

Budget the minutes and the files.

Provisional fares. These are the published limited beta fares. Check the fare table before submitting work.

The list fare is $0.00044 per audio minute, with a $0.0075 minimum per submission. A file under 30 seconds bills as 30 seconds. DeepInfra costs less on a short Whisper job. A completed transcription bills its units even when the transcript is empty.

Longer files can clear the submission minimum individually. In that case the same work pays the same list fare alone or in a batch. Combining files into one submission does not remove the 30-second minimum per file.

DeepInfra beats us on a short Whisper job. Our lone one-minute file costs $0.0075, with a $0.0075 minimum per submission. A file under 30 seconds bills as 30 seconds. DeepInfra lists $0.00045 per minute. Competitor prices checked 2 Nov 2026.

Download the subtitle files and review them.

Expect a cold start of 10 to 30 s, occasionally minutes. This is a queue: check the batch state instead of assuming a fixed completion time.

Run a dry run, review the hold and missing files, then submit with an idempotency key. Poll until settled_at is set and read succeeded items by page. Save each signed output URL as an .srt file, without sending your API key to storage.

Listen for missing speech, names, punctuation, and timing before publishing subtitles. Transcription does not add speaker labels or read text from video frames. Keep your own downloads: Deletion starts at 7 days and normally finishes within a further day.