What it does, and doesn't
Does
Transcribes recorded audio in ~100 languages, detects the language, adds punctuation, produces timestamps, translates non-English speech into English.
Doesn't
Label speakers, run in real time by default, or guarantee accuracy on noisy audio. It can invent text on silence or music.
Model sizes
| Model | Parameters | Accuracy | Speed | Typical use |
|---|---|---|---|---|
| tiny / base | 39M / 74M | Low | Very fast | Drafts, weak hardware |
| small | 244M | Good | Fast | Laptops without a GPU, clean English |
| medium | 769M | Very good | Moderate | CPU default for most languages |
| large-v3 | 1.55B | Best | Slow without GPU | Accents, jargon, noise; anything with a GPU |
Variants such as large-v3-turbo trade a little accuracy for much higher speed, and distilled versions exist for English. Apps usually let you switch.
Local Whisper vs the Whisper API
The same model, two very different privacy positions. Through the API, your audio is uploaded to OpenAI and billed per minute. Run locally, the audio never leaves your machine and there is no per-minute cost; you pay with compute time instead. A one-hour file takes a few minutes on a modern GPU and about an hour on a CPU.
- Command line: OpenAI's reference implementation, faster-whisper, whisper.cpp, WhisperX with speaker labels. Free, technical.
- Desktop apps: MacWhisper on Mac, Buzz on Mac and Windows, Scrieb on Mac and Windows with speaker identification and DOCX export.
Comparison of the apps: best local transcription software. What "local" has to mean: local transcription.
Accuracy in practice
On clear recordings of native speakers, large-v3 is close to human-level for most major languages. Accuracy drops with strong accents, domain jargon, crosstalk and low-quality audio. Where the model runs makes no difference: a cloud service using Whisper and a local app using the same model size produce the same text. What differs is who holds the audio.
Frequently asked questions
- Is Whisper free?
- The model is open source under the MIT licence and free to download and run. OpenAI also sells access to it through an API, priced per minute of audio, where your audio is uploaded to OpenAI. Running it locally costs nothing beyond your hardware.
- Does Whisper work offline?
- Yes, when you run it on your own computer through the command line or a desktop app. The API version requires an internet connection and uploads the audio.
- Which Whisper model should I use?
- large-v3 for the best accuracy if you have a GPU or patience; medium as the default on CPU-only machines; small or base for quick drafts of clean audio. Bigger models are more accurate and slower.
- Does Whisper identify speakers?
- No. Whisper only produces text. Speaker labels require a separate diarization model run alongside it, which some apps, Scrieb among them, include.
- What is Whisper hallucination?
- On silence, music or very poor audio, Whisper can invent plausible text that was never said, or repeat a phrase. It happens rarely on clear speech. Tools mitigate it with voice activity detection and by trimming silence; a read-through catches the rest.
- Can Whisper transcribe in real time?
- Not by design; it processes chunks of recorded audio. Streaming setups exist but add complexity and lag. For live captions, purpose-built streaming models are better; for recordings, Whisper is the standard.