Whisper

Our verdict

The reason per-minute transcription pricing is hard to defend. Whisper is MIT licensed, the weights are published, and it runs on your own machine, so recordings never leave it. OpenAI publishes a proper speed-and-memory ladder rather than marketing numbers. Two honest limits: accuracy varies a lot by language, and there is no speaker identification, so treat this as an engine to build on rather than a finished meeting tool.

Key features

  • Six model sizes from tiny (39M parameters) to large (1550M), plus an optimised turbo variant
  • English-only variants for four sizes, which OpenAI notes perform better at the small end
  • Multilingual speech recognition, speech translation and language identification in one model
  • Command-line tool and a Python API
  • Timestamped output suitable for subtitles
  • Runs entirely offline once the model is downloaded

Strengths

  • MIT licensed, read directly from the LICENSE file in OpenAI's own repository
  • No per-minute charge, and audio never leaves your machine - which matters for meeting recordings and client calls
  • A genuine speed and accuracy ladder: OpenAI publishes VRAM and relative speed for every model size
  • The turbo model is described as an optimised large-v3 with "minimal degradation in accuracy" at roughly eight times the speed
  • Broad multilingual coverage, with per-language error rates published in the paper

Limitations

  • Accuracy "varies widely depending on the language" in OpenAI's own words, and the published error rates differ sharply between them
  • The large model needs roughly 10 GB of VRAM, which rules out a lot of laptops
  • The turbo model "is not trained for translation tasks", so translating non-English speech needs one of the multilingual models instead
  • No speaker identification built in; separating who said what needs additional tooling
  • It is a model and a command-line tool, not a meeting product: no recording, no interface, no summaries

Who should skip Whisper

Anyone wanting a finished meeting-notes product with recording, speaker labels and summaries, since this is an engine to build on. Also anyone without a capable GPU who needs the larger models at speed.

Pricing details

Checked on the official pricing page on . Prices change often, so confirm before you buy.

What the free plan includes

  • There is no paid tier and nothing is withheld. The MIT licence covers the code, and every model size is downloadable. Your hardware is the limit: OpenAI's own table puts large at roughly 10 GB of VRAM, turbo at about 6 GB, medium at about 5 GB, and tiny at about 1 GB.

Worth knowing

Whisper has no pricing page because it has no price. The repository is MIT licensed and the model weights are published, so the cost is hardware and electricity rather than licensing. Worth separating from OpenAI's hosted audio transcription API, which bills per minute and is a different, paid product that happens to share the name. Checked 23 October 2026.

Commercial use

Allowed. Whisper is released under the MIT License, read directly from the LICENSE file in OpenAI's own repository, copyright dated 2022. MIT permits commercial use, modification, redistribution and sublicensing, including running it as a service for clients, provided the copyright notice and permission notice are retained. This is worth separating clearly from OpenAI's hosted audio transcription API, which is a paid product governed by OpenAI's commercial terms rather than by this licence. Using the open model locally and calling the hosted API are two different arrangements with two different sets of obligations. Checked 23 October 2026.

Notable integrations

  • Python API for use inside your own code
  • Command-line interface for batch transcription
  • Requires ffmpeg, available from most package managers
  • Model weights downloaded and cached locally
  • Widely re-implemented, including faster community ports

Alternatives to Whisper

  • Otter.ai

    A meeting assistant that joins calls on Zoom, Microsoft Teams and Google Meet, transcribes them as they happen, and leaves behind a searchable transcript with a summary and action items.

  • Fireflies.ai

    A meeting notetaker that records and transcribes calls, writes summaries, and exposes the results through an API, with transcription included on every tier including the free one.

  • Descript

    Descript is an AI-powered editor for turning recorded audio and video into polished content through text-based editing, transcription and collaborative workflows.

Guides