OpenAI Whisper review

What it is

Whisper is OpenAI's open-source speech recognition model. It is not an app — it is weights and code you run yourself, or a cheap hosted API endpoint if you would rather not.

Where it shines

Cost at scale. Transcribing ten thousand hours of audio through a commercial service is an enormous bill; through a local Whisper deployment it is electricity and time. Media companies, research groups and anyone with an archive use it for exactly this.

Privacy follows. Because it runs entirely on your machine, sensitive recordings — legal interviews, medical dictation, confidential meetings — never touch a third party. For regulated work that is often the only acceptable option.

Multilingual performance is the third strength. Ninety-plus languages with genuinely usable accuracy, and translation into English built in, at a quality level that many paid services do not match for non-English audio.

Community implementations have made it fast: optimised builds transcribe an hour of audio in a couple of minutes on ordinary hardware.

Where it falls short

There is no product. No interface, no editor, no sharing, no summaries — just a transcript file. Anyone who wants meeting notes, highlights or collaboration needs to build that or buy a tool that already has it.

Speaker diarisation is absent and must be added with separate libraries, which is fiddly. Long recordings need chunking, and the model can hallucinate plausible text during long silences or music, so pipelines need guards against that.

Pricing in plain English

Free to run locally under the MIT licence. About $0.006 per minute through the hosted API, and faster third-party deployments of the same weights can be even cheaper — all still well under most commercial transcription services.

Who should use it

Developers, researchers and organisations with high transcription volume or privacy requirements. Anyone who wants a polished experience should buy Otter, Fireflies or Descript instead.