Audio Transcriber Pro – Whisper large-v3-turbo (high accuracy)
Pricing
from $45.00 / 1,000 audio minute (pro / large-v3-turbo)s
Audio Transcriber Pro – Whisper large-v3-turbo (high accuracy)
High-accuracy Whisper large-v3-turbo transcription from direct URLs (CPU int8, beam 5, VAD). DE WER ≈5.5% FLEURS / ≈8.5% MLS (60 clips). Failed files never charged. No YouTube scraping.
Pricing
from $45.00 / 1,000 audio minute (pro / large-v3-turbo)s
Rating
0.0
(0)
Developer
Kilian Schieri
Maintained by CommunityActor stats
0
Bookmarked
2
Total users
1
Monthly active users
a day ago
Last modified
Categories
Security
Limited permissionsShare
Audio Transcriber Pro – Whisper large-v3-turbo
High-accuracy transcription of audio or video from direct URLs with faster-whisper large-v3-turbo on CPU (int8), beam size 5 and VAD. The model is baked into the image for faster cold starts.
Measured German WER (own bench, 2026-11-05): ≈ 5.5 % on FLEURS de_de (60 clips, 1375 words, int8); ≈ 8.5 % on MLS German (60 clips). Close to AssemblyAI Universal-3.5 Pro (~4.9 % FLEURS DE) and better than many Whisper-small Store actors.
No YouTube / TikTok / Instagram page URLs – pass a direct file link. Failed files are ERROR rows and never charged.
For a cheaper / faster option see datvanu/audio-transcriber (tiny/base/small).
Input
| Field | Default | Notes |
|---|---|---|
urls | – | Direct http(s) links, max 50 |
model | large-v3-turbo | Fixed |
language | auto | e.g. en, de |
task | transcribe | or translate |
includeSegments | true | Timestamped segments[] |
beamSize | 5 | Quality setting |
vadFilter | true | Skip long silence |
maxFileBytes | 200 MB | Oversized → ERROR, not charged |
Prefill: short public-domain Aesop fable (~46 s) from LibriVox / archive.org.
Output
One dataset item per URL: status, url, model, task, language, languageProbability, durationSecs, audioMinutes, text, segments[], error, transcribedAt.
Pricing (PPE)
| Event | When |
|---|---|
audio-minute-pro | started minute of audio (ceil) with large-v3-turbo |
apify-actor-start | Actor start (platform standard) |
Failed files are never charged. Remove synthetic apify-default-dataset-item in Console.
Memory
Default 4096 MB. Raise to 8 GB for very long files if needed.
Licence
- Engine: faster-whisper / CTranslate2, MIT.
- Prefill: LibriVox Aesop (public domain).
- You are responsible for the right to transcribe submitted files.
