Audio & Video Transcriber — Whisper (no API keys)

Transcribe audio and video files from URLs with OpenAI's Whisper model running inside the actor — no API keys, no external transcription service, nothing to configure. Paste direct file URLs, get back plain text, timestamped segments and ready-to-use SRT subtitles.

• Model: Whisper small (int8, CPU) via faster-whisper. The model is baked into the actor image, so runs start transcribing immediately — no download wait.
• Formats: anything FFmpeg can decode — mp3, wav, m4a, ogg, flac, opus, and the audio track of video files (mp4, webm, mov, mkv…).
• Languages: Whisper's ~100 languages, auto-detected by default (detected language + probability included in every record). Optional translate-to-English.

Input

{
  "audioUrls": ["https://raw.githubusercontent.com/lordbasilaiassistant-sudo/b2b-assets/main/gettysburg_address_64kb.mp3"],
  "language": "",
  "format": "both",
  "translateToEnglish": false,
  "maxFileSizeMb": 200
}

• audioUrls — direct URLs to audio/video files. One dataset record per URL.
• language — ISO 639-1 code (en, es, de, …); empty = auto-detect.
• format — both (default), text (plain text only) or segments (timestamped segments + SRT only).
• translateToEnglish — use Whisper's translate task instead of transcribing in the original language.
• maxFileSizeMb — larger downloads are skipped (recorded as failed, never charged).

Output

One dataset record per file. Real output from the default input above (a 2:38 public-domain LibriVox recording of the Gettysburg Address), text/segments/srt truncated here for space:

{
  "url": "https://raw.githubusercontent.com/lordbasilaiassistant-sudo/b2b-assets/main/gettysburg_address_64kb.mp3",
  "ok": true,
  "durationSeconds": 157.88,
  "language": "en",
  "languageProbability": 0.9997,
  "wordCount": 299,
  "tookMs": 315181,
  "text": "This is a LibriVox recording. All LibriVox recordings are in the public domain. [...] Four score and seven years ago, our fathers brought forth on this continent a new nation conceived in liberty [...]",
  "segments": [
    { "start": 0.05, "end": 6.01, "text": "This is a LibriVox recording. All LibriVox recordings are in the public domain. For more" },
    { "start": 6.01, "end": 13.13, "text": "information or to volunteer, please visit LibriVox.org." }
  ],
  "srt": "1\n00:00:00,050 --> 00:00:06,010\nThis is a LibriVox recording. [...]"
}

Failed files push { "url": …, "ok": false, "error": … } and are never charged.

Speed and limits (measured, not estimated)

• Runs Whisper small on CPU — solid accuracy for clear speech; it is not a GPU service and won't match large-model accuracy on very noisy audio or heavy accents.
• Measured on the default 2 GB actor memory (0.5 CPU core), three runs of the 2:38 test file: 4.2, 5.3 and 7.2 minutes wall — about 1.6 to 2.7 minutes of processing per minute of audio depending on platform CPU contention. CPU share scales with actor memory, so raising memory to 4 GB (1 full core) roughly halves it. Memory peak measured: 1.24 GB.
• Download timeout 30 s per file; size cap 200 MB (configurable down).
• Segments are phrase-level (start/end seconds); word-level timestamps are not included.

Billing and price comparison

$0.03 per started minute of audio actually transcribed (one audio-minute event per started minute — the 2:38 example = 3 events, $0.09). Failed downloads, oversized files and undecodable inputs are recorded in the dataset but never charged.

Measured against incumbents (store prices, 2026-08-07):

Actor	Price
This actor	$0.03 / audio-minute
tictechid/vanzi-universal-transcriber	$0.0025 / transcription-second = $0.15 / audio-minute
memo23/video-audio-transcriber	$0.02–0.05 / audio-minute

Honest note on speed: processing is CPU Whisper small (typically 0.4–0.6x realtime at default memory; measured 0.36–0.63x across our runs) — cheaper than GPU services because you're not paying for idle GPU time, at the cost of slower turnaround on long files.