What it is
Whisper large-v3-turbo is a pruned and fine-tuned Whisper large-v3, released by OpenAI in October 2024 under the MIT licence. It keeps the large-v3 encoder but cuts the decoder from 32 layers to 4, which OpenAI's GitHub README puts at about eight times the speed of large-v3 in roughly 6GB of VRAM. Its Hugging Face page records more than 124 million downloads.
OpenAI notified developers on August 26, 2026 that its hosted whisper-1 API model and the /v1/audio/translations endpoint will be removed on February 26, 2027, in favour of gpt-transcribe and gpt-live-transcribe. The open weights are unaffected, but OpenAI has not shipped a new Whisper checkpoint since this one.
Our review
Whisper large-v3-turbo survives on price and licence, not accuracy. Artificial Analysis measures it at 4.6% AA-WER on Groq, at $0.67 per 1,000 minutes and about 200 times real-time. ElevenLabs Scribe v2 scores 2.2% and Mistral's Voxtral Small Transcribe 3 scores 2.1%, so the leaders make fewer than half its errors, at five to six times the price.
The speed comes from a blunt cut: four decoder layers instead of 32. OpenAI's README is explicit that turbo was not trained for translation, so anyone who relies on Whisper's translate-to-English mode needs large-v3 instead, and Groq's own docs list translation as unsupported on turbo. Groq also publishes its own word error rate for the hosted model, 12% against 10.3% for large-v3, which is a useful reminder that turbo trades some accuracy for throughput on every host.
The model card's warning about hallucination still applies. Whisper was trained on weakly labelled web audio and can emit text that was never spoken, and OpenAI reports uneven accuracy across languages, accents and dialects, with lower accuracy on low-resource languages.
What keeps it in use is everything built around it: faster-whisper, whisper.cpp, Transformers pipelines, countless fine-tunes and hosts from Groq to Together AI. Nvidia's Parakeet TDT 0.6B v3 now scores a slightly better 4.5% on the same leaderboard and runs faster on Together AI, but at $1.50 per 1,000 minutes it costs more than twice as much as Groq's turbo.
Where it earns its place
A podcast network with 10,000 hours of back catalogue
Make every episode searchable and generate show-note drafts
Batch the archive through Groq at $0.04 an hour, about $400 for the lot, then send only flagged episodes to a pricier model for correction. Accuracy is good enough for search, not for verbatim quotes.
A platform engineer at a health-tech start-up that cannot send audio to third parties
Transcribe clinician dictation on premises
Run turbo through faster-whisper on one 8GB or larger GPU inside the private network. There is no API bill or vendor data-processing agreement, but the team owns evaluation and error correction for medical terms.
A two-person team building a voice-notes app
Add on-device transcription without a per-minute bill
Ship whisper.cpp with the turbo weights; the MIT licence allows bundling in a paid app with no attribution beyond the licence notice. Expect weaker results on accents and low-resource languages.
Before you commit
- Setup
- pip install openai-whisper and run whisper audio.mp3 --model turbo, or call Groq's OpenAI-compatible transcription endpoint with whisper-large-v3-turbo; either takes minutes.
- Integrations
- Transformers, faster-whisper, whisper.cpp and most speech toolkits; hosted by Groq and Telnyx among others, with Groq accepting files up to 25MB on the free tier and 100MB on the dev tier.
- Data and privacy
- Open weights under MIT, so self-hosting keeps audio entirely in your control; on hosted APIs, retention follows the host's terms rather than OpenAI's.
- Cost at scale
- Groq bills $0.04 per audio hour with a 10-second minimum per request, so 100,000 hours a month is about $4,000; self-hosting needs only a GPU with roughly 6GB of VRAM.
- Limits
- 30-second receptive field, so long audio is chunked or processed sequentially; no translation; Groq quotes 12% WER for its turbo endpoint against 10.3% for large-v3.
What's good
- MIT licence with no field-of-use or revenue limits
- $0.04 per audio hour on Groq
- Runs in about 6GB of VRAM, so a single consumer GPU is enough
- Supported by almost every speech toolkit and host
What's not
- 4.6% AA-WER, roughly double the errors of Scribe v2 or Voxtral Transcribe 3
- Not trained for translation, unlike large-v3
- Can hallucinate text that was never spoken
- No new Whisper checkpoint since October 2024
Best for
Teams that need cheap, permissively licensed transcription they can self-host or buy for cents an hour, and can live with second-tier accuracy.
Alternatives
Whisper large-v3-turbo: common questions
- Is Whisper large-v3-turbo worth it?
- Whisper turbo is now the cheap, unrestricted baseline that newer speech models are measured against, not the model that wins. We rate it 3.7 out of 5.
- How much does Whisper large-v3-turbo cost?
- Free under the MIT licence. Hosted versions are cheap: Artificial Analysis lists Groq at about $0.67 per 1,000 minutes.
- What are the downsides of Whisper large-v3-turbo?
- 4.6% AA-WER, roughly double the errors of Scribe v2 or Voxtral Transcribe 3; not trained for translation, unlike large-v3; can hallucinate text that was never spoken; no new Whisper checkpoint since October 2024.
- Who is Whisper large-v3-turbo best for?
- Teams that need cheap, permissively licensed transcription they can self-host or buy for cents an hour, and can live with second-tier accuracy.
- How much does Whisper large-v3-turbo cost for a team or in production?
- Groq bills $0.04 per audio hour with a 10-second minimum per request, so 100,000 hours a month is about $4,000; self-hosting needs only a GPU with roughly 6GB of VRAM.
- Is Whisper large-v3-turbo safe for company data?
- Open weights under MIT, so self-hosting keeps audio entirely in your control; on hosted APIs, retention follows the host's terms rather than OpenAI's.
- What are the best Whisper large-v3-turbo alternatives?
- The closest alternatives we have reviewed are ElevenLabs Scribe v2, GPT Transcribe, Deepgram Nova-3 and Parakeet TDT 0.6B v3.
Answers checked Oct 22, 2026.