Whisper large-v3-turbo 🧠 Model Open source

What it is

Whisper is OpenAI's open-source speech recognition model, released under the MIT licence. It transcribes and translates speech in 99 languages and detects which language is spoken. The newest open checkpoint is turbo (large-v3-turbo). It is an 809M-parameter pruned version of large-v3, about 8x faster with little loss in accuracy. OpenAI's paid API still offers whisper-1, but it now points users to newer models (gpt-4o-transcribe, gpt-realtime-whisper).

Key features

• 99-language transcription, translation to English and language detection
• Six sizes from tiny (39M) to large (1.55B). Turbo needs about 6 GB of VRAM
• Timestamps for subtitle (SRT/VTT) generation
• Runs fully offline on your own machine
• The base for many community ports (whisper.cpp, faster-whisper) and desktop apps

Where you can use it

Install it with pip and run it locally, or call whisper-1 through the OpenAI API. Many transcription and captioning tools build on Whisper, but few name it on their pricing or product pages, so we do not link specific apps here.

Pricing and rights

The open-source model is free under the MIT licence, including commercial use. You pay only for your own hardware. The hosted whisper-1 API costs $0.006 per minute. Accuracy drops on noisy audio and rare languages, and Whisper can produce made-up text during silence. Proofread before you publish subtitles.

Who it is for

Podcasters, video editors and developers who want free, private, offline transcripts and subtitles, or a base for their own captioning tool.

Verdict

Whisper is still the default open speech-to-text model: free, multilingual and easy to run. It has had no major new open checkpoint since turbo. It does not separate speakers on its own, and hallucinated text on silent passages means you still need a human pass.