Transcription · OpenAI (Wizper: fal)
Whisper
OpenAI's open speech model: the widest language coverage here at the same low price as Parakeet, in private mode.
Private Low costWhat sets it apart
Whisper is OpenAI's open speech model, trained on 5 million hours of audio and covering 99 languages, the widest range here. Large v3 cut errors by 10% to 20% compared with large v2, and Wizper serves the same weights faster.
- 99 languages, Apache 2.0
- 1.55B parameters, large v3
- Wizper: same weights, served faster
Recommended for: transcribe several languages or strong accents.
Strengths and limits
Where it shines
- Best multilingual accuracy in Sub Rosa's suggestions
- Close to 100 languages
- Same low price as Parakeet, private mode
Think twice when
- Can invent text over long silences or noise
- Whisper Large V3, open weights
Use cases
Transcribe an interview in a less common language
Transcribe this interview recorded in Polish, then give me an English translation below it.
It covers 99 languages, more than any other option here.
Get a transcript back quickly
Transcribe this 20-minute podcast episode as plain text, no timestamps.
Wizper returns the same transcript faster, per fal.
Against its rivals
- ParakeetParakeet is smaller and punctuates by default, for 25 languages only.
- ScribeScribe adds speaker labels and sound tags, is more accurate on the Artificial Analysis test, and costs more.#1 of 4 on Artificial Analysis word error rate (2.2%)
- xAI Speech to TextxAI is cheaper and more accurate on that test, with 25+ languages.#2 of 4 on Artificial Analysis word error rate (4%)
Version history
Each version, dated, with what it changed over the one before.
- Whisper large v2Earlier, not in Sub Rosa
- Open multilingual speech model with 1.55B parameters
- Whisper large v3
- 128 mel frequency bins instead of 80
- New Cantonese language token and 10% to 20% fewer errors than large v2
- Wizper (Whisper v3, fal edition)Latest
- Whisper v3 large with optimized serving: same word error rate, twice the speed, per fal
Benchmarks
How to read these scores: what a benchmark tells you.
Artificial Analysis word error rate: #3 of 4 models measured in the catalog. The share of words a transcription gets wrong (missed, added or replaced) on a mixed set of real recordings. Lower is better. At 5%, one word in twenty needs fixing; under 3% a transcript reads almost clean.
See the numbers
| Model | % |
|---|---|
| 1. ElevenLabs Scribe V2 | 2.2% |
| 2. xAI Speech to Text v1 | 4% |
| 3. Whisper Large V3 | 4.1% |
| 4. Wizper (Whisper v3) | 4.7% |
All 2 models in this family
| Model | Use | Privacy | Price |
|---|---|---|---|
Whisper Large V3openai/whisper-large-v3 | Transcription | Private | $0.162 per hour of audio |
Wizper (Whisper v3)fal-ai/wizper | Transcription | Private | $0.162 per hour of audio |
Specifications
| Specification | Whisper large v3 |
|---|---|
| Released | November 2023 |
| Parameters | 1.55B |
| Weights | Open, Apache 2.0 |
| Accepts | audio |
| Produces | text |
| Languages | 99 languages |
| Price | $0.162 per hour of audio |
| Privacy | Private |
Prices, context and lengths from the live catalog on October 16, 2026; parameters and licenses from the maker.
Sources
Everything on this page comes from these pages, read on October 16, 2026 or as dated on each score.