Media Transcriber: Audio & Video to Text, SRT and VTT (Whisper)

By Blackcat Forge

Turn your own audio and video files into clean text, timestamped segments, SRT and WebVTT subtitles in bulk. Paste direct file links (MP3, WAV, M4A, AAC, OGG, FLAC, MP4, MOV, WebM, MKV…) and get one result row per file, plus downloadable .txt, .srt and .vtt files.

Good for podcast show notes, subtitle files for your own videos, searchable meeting and lecture archives, interview transcripts, and feeding speech into LLM pipelines.

Why this one

• Pay per audio minute, nothing else. No monthly fee, no per-file minimum beyond one minute. Files that fail are free.
• Three engines:
  • Fast (Whisper base, or tiny for maximum speed): roughly 10–30× faster than real time on the default 4 GB setup.
  • Accurate (Whisper small): better with accents, noisy audio and non-English speech.
  • Bring your own key (Groq whisper-large-v3-turbo or OpenAI whisper-1): best accuracy and speed. You pay the provider directly, and this Actor charges a tiny orchestration fee.
• 99 languages with auto-detection, or translate any language to English.
• Subtitles ready to upload: SRT and VTT with correct timestamps, plus optional word-level timing.
• Long files are fine: audio is processed in 10-minute windows, so memory stays flat for multi-hour recordings.
• Private by design: files are deleted from the run's disk right after processing. Your API key is stored encrypted by Apify and is never logged or written to results.

Input

{
  "mediaUrls": ["https://example-bucket.s3.amazonaws.com/episode-42.mp3"],
  "model": "base",
  "language": "",
  "task": "transcribe",
  "outputFormats": ["text", "segments", "srt", "vtt"],
  "wordTimestamps": false
}

Links must be direct links to a file that you own or have the rights to transcribe. That includes your own storage bucket or CDN, a podcast RSS enclosure, or a Dropbox/Google Drive direct-download link. Links to YouTube, TikTok, Instagram, Facebook, X, Vimeo, Spotify, SoundCloud and similar platforms are refused, because their terms don't allow downloading.

Output (one row per file)

{
  "url": "https://…/episode-42.mp3",
  "status": "ok",
  "durationSeconds": 506.9,
  "billedMinutes": 9,
  "language": "en",
  "engine": "faster-whisper/base",
  "wordCount": 1312,
  "text": "This is a LibriVox recording…",
  "segments": [{ "start": 0.0, "end": 4.2, "text": "This is a LibriVox recording." }],
  "srt": "1\n00:00:00,000 --> 00:00:04,200\nThis is a LibriVox recording.\n…",
  "files": { "txt": "https://api.apify.com/v2/key-value-stores/…/records/001-episode-42-….txt", "srt": "…", "vtt": "…" }
}

Failed files come back with status: "error" and a plain-English error, and they're not charged.

Pricing (pay per event)

Event	Price
Audio minute, fast model (tiny / base)	$0.005 per started minute
Audio minute, accurate model (small)	$0.015 per started minute
Audio minute, your own Groq/OpenAI key	$0.002 per started minute (+ what your provider charges you)
Large file download, per 100 MB beyond 3 MB per audio minute	$0.03
Actor start (covers loading the speech model)	$0.0005 per GB of run memory ($0.002 per run at the default 4 GB)

Examples: a 45-minute podcast on the fast model costs $0.23 (plus the $0.002 start fee). Ten hours of lectures cost $3.00. Batch many files into one run to pay the start fee only once.

The large-file fee only applies to heavy files, such as a high-bitrate 1080p video where you only need the speech. Audio files of any normal quality, even 320 kbps, never hit it. Tip: send audio-only files when you can.

Set Max cost per run and the Actor stops cleanly at your limit. If the limit is reached partway through a file, the file is transcribed only up to the minutes you can afford, and that's all you're charged for.

Accuracy notes

• base is good for clear speech: podcasts, narration, meetings with decent mics. small is noticeably better for accents, crosstalk, noise and most non-English languages. For the best quality, bring a Groq key: large-v3-turbo costs about $0.04 per hour at Groq, so $0.002 + ~$0.0007 per minute in total.
• Silence and music-only parts are skipped (voice-activity detection), so they don't produce hallucinated text.
• Whisper doesn't label speakers (no diarization).

Limits

• Up to 600 minutes per file (default 240) and 4 GB per file (default 1 GB).
• The fast and accurate models run on CPU. Very long files take a while, roughly 4–6 minutes per hour of audio on the fast model and 15–25 minutes on the accurate one. Word-level timestamps add about 70%.

Responsible use

Only transcribe recordings that you own or are allowed to process, and follow the recording-consent laws that apply to you. This Actor does no scraping: it only downloads the exact file URLs you provide. It refuses private-network addresses and platforms whose terms forbid downloading.

Local development

pip install -r requirements.txt     # also needs ffmpeg on PATH
python -m pytest tests
apify run                            # reads storage/key_value_stores/default/INPUT.json

Blackcat Forge builds small, dependable tools for developers and creators.