Audio & Video File Transcriber – Whisper to Text, SRT, VTT

Turn direct links to audio or video files you have the right to use (mp3, wav, m4a, mp4, ogg, webm, flac …) into:

• plain text,
• timestamped segments (start, end, text),
• SRT and WebVTT subtitles (in the dataset item and as files in the run's key-value store).

It runs the open-source Whisper model (MIT license) through faster-whisper on CPU inside the Actor. No API key and no third-party transcription service is needed.

Unofficial. This Actor is not affiliated with, endorsed by or connected to OpenAI, YouTube, TikTok, or other companies or platforms.

Pricing

Pay-per-event:

• US$0.005 per run (start event, charged once per run, not per file, and only if at least one file is transcribed), plus
• US$0.004 per audio minute.

Put many files into one run. The start fee and the model start-up time are paid once per run, so 20 files in one run cost US$0.005 in start fees, while 20 separate runs cost US$0.10. fileUrls takes a list. Duplicate links are removed before anything is downloaded or charged. Matching ignores #fragments, tracking parameters (utm_*, fbclid, gclid and similar) and letter case in the domain. A duplicate gets a free row with status duplicate.

Example: one run with a 161 s file and a 263 s file = 0.005 + (3 + 5) × 0.004 = US$0.037.

• Each successfully transcribed file is charged for its duration rounded up to the next whole minute (24 s = 1 minute, 161 s = 3 minutes). The chargedMinutes field shows what each file cost.
• Failed downloads, rejected URLs, unreadable files and files skipped because of your spending limit are free. If no file in the run is transcribed (for example, because every link is invalid or rejected), the run costs nothing: the start fee is charged only together with the first transcribed file.
• Files with no speech are free. If no speech is detected in a file (for example only music or silence), it gets status no-speech, is not charged, and does not count as a transcribed file, so on its own it does not trigger the run fee. With vadFilter: false, Whisper may mistake music for speech and output text for it; such a file gets status ok and is charged like any other transcribed file.
• Your maximum total charge (maxTotalChargeUsd) includes the start fee. If it cannot cover the start fee plus one audio minute (US$0.009), the run stops without charging anything. A file that would exceed the remaining budget is skipped, not charged.
• A restarted or migrated run does not charge the start fee or the same file a second time.

Accuracy (internal test, 2026-10-17, prototype version run locally; not promised)

• English: word error rate 4.6–7.8% with the default base model on 4 public-domain LibriVox recordings (2.7–18 min), measured after trimming the LibriVox intro/outro. Clear read speech; noisy audio, overlapping speakers, strong accents and music will be worse.
• Chinese: output may mix Simplified and Traditional characters in the same transcript (see the example below). Convert afterwards (e.g. OpenCC) if you need one script.
• Names and rare words are often misspelled (e.g. "LibriVox" → "Libravox").
• tiny is faster but less accurate than base.

Input

Field	Default	Notes
fileUrls	–	Direct file links. Platform page URLs (YouTube, TikTok, Instagram, Facebook, X, Vimeo, SoundCloud, Spotify, Bilibili …) are rejected: this Actor does not download from such platforms.
model	base	base (more accurate) or tiny (faster).
language	auto	ISO code (en, zh, es …) to skip auto-detection.
task	transcribe	translate = English output for non-English audio.
vadFilter	true	Skips silence/music; reduces hallucinations. With false, music may be mistaken for speech and charged.
beamSize	1	1 = greedy (fastest).
maxFileSizeMb	200	Hard limit (max 500).
maxDurationMinutes	120	Hard limit per file (max 120). Longer files are rejected and not charged.

Each file gives one dataset item with status = ok, no-speech, error or duplicate, and a readable error message when it is not ok. no-speech means the file was processed but no speech was detected; it is free. duplicate means the link points to the same file as an earlier entry in the same run, after ignoring #fragments, tracking parameters (utm_*, fbclid, gclid and similar), letter case in the domain and the default port. A duplicate is not downloaded or transcribed, is free, and its error text names the earlier entry. This applies even if the earlier entry failed. One bad URL does not fail the whole run. If the file host answers HTTP 429/5xx, the Actor waits (honouring Retry-After) and retries up to 3 times, then reports the file as failed (free). It does not use proxies or try to get around rate limits.

Output (real rows from a platform run, text shortened)

{"url": "https://archive.org/download/gettysburg_address_librivox/gettysburg_address_lincoln_64kb.mp3", "status": "ok", "language": "en", "durationSeconds": 161.18, "chargedMinutes": 3, "segmentCount": 29, "text": "This is a Libravox Recording. All Libravox recordings are in the public domain. …", "srt": "1\n00:00:01,900 --> 00:00:04,620\nThis is a Libravox Recording.\n\n2\n…"}
{"url": "https://archive.org/download/call_to_arms_jl_librivox/calltoarms_06_lu_64kb.mp3", "status": "ok", "language": "zh", "durationSeconds": 262.95, "chargedMinutes": 5, "segmentCount": 45, "text": "魯迅小说急,那喊,一见小事。此次Librivalc's錄音由公眾所有。我从相下跑到京城里,一转眼已经六年了。…"}
{"url": "https://www.youtube.com/watch?v=dQw4w9WgXcQ", "status": "error", "error": "Platform URLs are not supported (youtube.com). …"}

Items also contain segments, vtt, processingSeconds, realTimeFactor, fileBytes and files (links to the .txt, .srt, .vtt records in the key-value store).

Speed

On the platform at 4 GB memory, transcription with base ran at about 0.08–0.12 × real time, plus about 10–20 s start-up per run. In internal tests (2026-10-17) a 4.4 min file took 30 s, an 8.5 min file took 50 s, and a 61.5 min audiobook chapter took about 7 min. Files longer than 15 minutes are transcribed in windows of about 10 minutes, cut at pauses. This keeps memory use roughly flat, so files up to the 120-minute limit fit in the default 4 GB. The default run timeout is 3 hours, enough for about 15 hours of audio in one run.

API examples (all three were run and returned data in an internal test, 2026-10-17, build 1.0.2)

curl

curl -s -X POST "https://api.apify.com/v2/acts/raspberry_rondel~audio-video-file-transcriber/run-sync-get-dataset-items?fields=url,status,language,durationSeconds,chargedMinutes,text" \
  -H "Authorization: Bearer $APIFY_TOKEN" -H "Content-Type: application/json" \
  -d '{"fileUrls": ["https://archive.org/download/sammlung_deutscher_gedichte_018_1506_librivox/deutschegedichte018_03_wunder_hn_64kb.mp3"], "model": "tiny"}'

This one-file example costs US$0.005 + 1 × US$0.004 = US$0.009. For real work, list all your files in fileUrls in one run. Python (pip install apify-client, v3)

import os
from apify_client import ApifyClient

client = ApifyClient(os.environ["APIFY_TOKEN"])
run = client.actor("raspberry_rondel/audio-video-file-transcriber").call(
    run_input={"fileUrls": ["https://archive.org/download/sammlung_deutscher_gedichte_018_1506_librivox/deutschegedichte018_03_wunder_hn_64kb.mp3"], "model": "tiny"}
)
for item in client.dataset(run.default_dataset_id).iterate_items():
    print(item["status"], item.get("language"), item.get("durationSeconds"), item.get("chargedMinutes"))
    print(item.get("srt", "")[:200])

JavaScript (npm install apify-client)

import { ApifyClient } from 'apify-client';

const client = new ApifyClient({ token: process.env.APIFY_TOKEN });
const run = await client.actor('raspberry_rondel/audio-video-file-transcriber').call({
    fileUrls: ['https://archive.org/download/sammlung_deutscher_gedichte_018_1506_librivox/deutschegedichte018_03_wunder_hn_64kb.mp3'],
    model: 'tiny',
});
const { items } = await client.dataset(run.defaultDatasetId).listItems();
for (const item of items) console.log(item.status, item.language, item.durationSeconds, item.chargedMinutes, item.text?.slice(0, 120));

Your responsibility

Only submit files you own or are allowed to transcribe. Transcripts can contain personal data spoken in the audio; you are responsible for handling it lawfully (e.g. under the GDPR or Taiwan's Personal Data Protection Act). Files are downloaded into the run's temporary storage and deleted after transcription; transcripts stay in your run storage under Apify's normal retention.

Copyright. Speech, music, lectures, podcasts and videos are usually protected by copyright. A transcript is a copy of the spoken words, and a translation (task: translate) is a derivative work (in Taiwan, translation is 改作 under Copyright Act Art. 3(1)(11) and Art. 28). Transcribe or translate only content you own, content you are licensed to use, or content where an exception such as fair use (US 17 U.S.C. §107) or 合理使用 (Taiwan Copyright Act Art. 65) clearly applies.

Why platform links are rejected. Platforms such as YouTube do not allow downloading except as they expressly authorise. YouTube's Terms of Service say you may not "access, reproduce, download … any Content except: (a) as expressly authorized by the Service; or (b) with prior written permission from YouTube and, if applicable, the respective rights holders". This Actor therefore accepts only direct file links and rejects platform URLs, including after redirects. Please do not use converter or mirror links to get around this.

Rights holders. If you believe this Actor has been used to process your content without permission, please contact us through this Actor's Issues tab on Apify.

This is general information, not legal advice.

Other Actors by the same developer

• Google Trends Scraper: interest over time, related queries and regions.
• Google Autocomplete Keyword Expander: A–Z and question keyword ideas.