OpenAI Whisper is a speech recognition model that turns spoken audio into written text in many languages. OpenAI released it as open source in September 2022, so you can download the model and run it on your own computer for free. OpenAI also sells transcription through its paid API, which charges your OpenAI account each time you use it.
"Whisper" can therefore mean two things: a free model you run locally, or a paid cloud service. Cost and privacy depend on which one you use.
What OpenAI Whisper is
Whisper is a model, not an app. You give it an audio clip and it returns text with punctuation and capitalization. It can also translate speech in other languages into English text. To use it, you need software that records or loads the audio, runs the model and shows you the result.
Four properties affect which app you should pick:
- It works on 30-second windows. The model processes audio in 30-second chunks. Apps that handle long recordings split the audio and join the text back together. If an app doesn't do this, long recordings get cut off.
- It is not live by design. Whisper transcribes a clip after you record it. Apps that look "live" feed it short chunks while you speak.
- It can invent text. On silence or heavy background noise, Whisper sometimes outputs words nobody said. Short, clear recordings give the cleanest results.
- It does not type anywhere on its own. The app around the model decides where the text goes, whether that's a note, a document or any text field.
The original release is a Python package from OpenAI (github.com/openai/whisper). Another common option is whisper.cpp, a C/C++ port that runs on ordinary computers without Python and uses compact model files in the "ggml" format. Gubb, our app, uses whisper.cpp.
Is Whisper free to use?
The model is free. Using it may or may not be, depending on how you run it.
| How you use Whisper | Cost | What you pay with |
|---|---|---|
| Open-source model (Python or whisper.cpp) on your own computer | Free (MIT license) | Disk space, RAM, processing time and setup effort |
| An app that bundles the model | Depends on the app: free tier, one-time price or subscription | The app's price. Transcription runs on your computer |
| OpenAI's transcription API | Paid per use, billed to your OpenAI API key | Money and an internet connection. Your audio goes to OpenAI |
Running Whisper locally never produces a bill, however much you transcribe. The API is pay-as-you-go. For per-minute rates and what an hour of audio costs, see our breakdown of OpenAI Whisper API cost.
Whisper model sizes and accuracy
Whisper comes in five sizes: tiny, base, small, medium and large. The smaller sizes also have English-only versions with a .en suffix. The large model has had several releases, including large-v2, large-v3 and a faster "turbo" variant.
Bigger models are more accurate, especially with accents, jargon and noisy audio. They also take more disk space, more memory and more time per recording.
These are the whisper.cpp models Gubb offers, with the sizes and RAM guidance the app shows:
| Model in Gubb | File | Size | RAM guidance |
|---|---|---|---|
| Whisper Small | ggml-small.bin |
480 MB | No specific requirement listed |
| Whisper Medium | ggml-medium.bin |
1500 MB | 8 GB or more recommended |
| Whisper Large | ggml-large-v2.bin (large-v2) |
3000 MB | 16 GB or more recommended |
We deliberately don't offer tiny and base.
A practical way to choose:
- Start with Small. It is the quickest download and needs the least memory.
- Move to Medium if Small keeps getting names, technical terms or accented speech wrong, and your Mac has at least 8 GB of RAM.
- Use Large only if you have 16 GB of RAM or more and accuracy matters more to you than waiting longer for each transcript.
Check your disk space as well. Keeping all three models takes about 5 GB.
Running Whisper locally vs using the API
| Local (on your computer) | OpenAI API | |
|---|---|---|
| Where your audio goes | Stays on your computer | Uploaded to OpenAI |
| Internet | Only to download the model once | Needed for every transcription |
| Cost | Free after setup | Billed per use to your API key |
| Speed | Depends on your hardware and model size | Depends on your connection and OpenAI's servers |
| Setup | Download a model file (an app can do this for you) | Create an OpenAI account and API key, add billing |
| Models | Whichever size fits your RAM | OpenAI's hosted models, such as whisper-1 and the newer GPT-4o transcribe models |
Pick local if:
- your recordings are private, such as confidential notes, client calls, or health or legal details,
- you often work offline, or
- you transcribe a lot and don't want a running bill.
Pick the API if your computer is old or low on memory, or if you don't want to download gigabytes of model files. Keep the API key private, because anyone who has it can run up charges on your account.
Hardware matters for local use. Superwhisper says its offline models run really well only on Apple silicon Macs, and that Intel Macs work best with its cloud models. On any slower machine, bigger models take longer to transcribe.
To run the open-source model yourself without writing code, see our guide to using Whisper on Mac without coding.
Apps that use Whisper on Mac
We make Gubb, one of the apps below, so weigh our view accordingly. The other two are strong choices for different jobs.
MacWhisper
MacWhisper transcribes audio and video files on your Mac using Whisper and other speech models. It also does dictation. Version 15 added real-time meeting transcription with speaker identification, which is in beta.
- Good at: transcribing recordings, interviews and videos. The free tier transcribes files with local models and exports TXT, SRT and VTT.
- Lacks: notes, lists or other tools. It is Mac only, and recent versions need macOS 15 or later. Batch transcription and speaker recognition need Pro.
- Price (as of November 2026): free tier available. Reviews put the direct Pro license at €64 one-time, though sources disagree. The App Store version ("Whisper Transcription") is priced separately. Check macwhisper.com before buying.
Superwhisper
Superwhisper is a dictation app that types into any app, using offline or cloud transcription. It runs on macOS, Windows, iOS and Android.
- Good at: dictating straight into email, chat or documents. It supports 100+ languages and custom AI modes.
- Lacks: a low-cost one-time option, since the lifetime license is $249.99. The Mac app needs macOS 14 or later, and offline models work best on Apple silicon.
- Price (as of November 2026): one review describes a free plan with small local Whisper models. Third-party pricing pages list Pro at $8.49/month, $84.99/year or $249.99 for a lifetime license. See superwhisper.com.
For a closer look at these two, see Superwhisper vs MacWhisper.
Gubb
Gubb is a scratchpad app for notes, checklists, clipboard history and timers, with speech-to-text built in. It runs Whisper on your Mac's CPU through whisper.cpp. You download the Small, Medium or Large model inside the app under Settings ▸ Voice to Text ▸ Available Models, and Gubb fetches it from the whisper.cpp project on Hugging Face. What to know before you pick it:
- Text goes into a Gubb note at the cursor. It does not type into other apps. There is no keyboard shortcut to start recording, so you click the mic.
- Local mode is English-only and transcribes only the first 30 seconds of each recording. Whisper supports other languages, but Gubb fixes the language to English. That makes it suited to quick thoughts and short notes, not long dictation.
- After downloading a model, select it under Active Model (or restart Gubb) before you record. Otherwise transcription fails.
- Gubb doesn't remember your model choice. At each launch it makes the first downloaded model active, checking Small, then Medium, then Large. If you keep Small installed but want Medium, reselect Medium after each restart.
- Gubb's settings also list an OpenAI API option, but in the current version it reports "OpenAI API is not configured" even after you save a key. Use the local models for now.
As of November 2026, Gubb costs $59 one-time from gubb.net, with a 7-day free trial of the direct download. It is also on the Mac App Store. For the full workflow, see how to dictate notes on Mac.
Which one fits
- Pick MacWhisper if you mainly transcribe files, such as recordings, podcasts, videos or meetings.
- Pick Superwhisper if you want Whisper dictation in every app, or in languages other than English.
- Pick Gubb if you want short English voice notes to land in a quick notes window you also use for lists, timers and clipboard history, and you prefer a one-time price.
- Use Apple Dictation if you just want free dictation in any text field. It isn't Whisper, but on Apple silicon Macs it runs on-device for supported languages. See how to use dictation on Mac.
FAQ
Does Whisper work offline?
The open-source model does. Once the model file is downloaded, local apps transcribe with no internet connection. The OpenAI API always needs a connection.
Do I need a GPU to run Whisper?
No. whisper.cpp runs on a regular CPU, and that is how Gubb runs it. Bigger models take longer on slower machines.
Is Whisper the same as the OpenAI API?
No. The API is OpenAI's paid hosted service, which offers whisper-1 and newer transcription models. The open-source Whisper model is a separate, free download that you run yourself.