What Whisper is

Whisper is an open speech recognition model released by OpenAI in September 2022, trained on 680,000 hours of multilingual audio and distributed in several sizes, from tiny (39 million parameters) to large (1.55 billion). On a Mac it runs locally through ports like whisper.cpp or WhisperKit, which load the model into memory and run it on the CPU, the GPU or the Neural Engine. The model is a file you download, and the app that uses it decides which size and which port.

Because it is open, dozens of dictation apps are built on it, and they compete on what they add around it: the interface, the modes, the cleanup, the choice of model.

What SpeechAnalyzer is

SpeechAnalyzer is the speech recognition engine Apple introduced with macOS 26, the successor to the older SFSpeechRecognizer API. It runs on the device, on the Neural Engine, with models that macOS downloads and manages itself, per language, the same way it manages keyboard dictionaries. Notes, Voice Memos and the system dictation use it, and third-party apps get it through a public framework. What SpeechAnalyzer changes, in detail.

Nothing about it is open: you cannot pick a model size or swap it for another. In exchange, nothing about it is yours to maintain: the model is updated with the operating system.

What changes for the user

 Whisper appsApple's engine (SpeechAnalyzer)
Where it runsYour MacYour Mac, on the Neural Engine
What you downloadA model file per size: 75 MB (tiny) to about 3 GB (large)Nothing; macOS manages the language models
Memory while runningThe model, loaded: from about 1 GB for tiny to about 10 GB for large, per OpenAI's tableIn system storage, outside the app's memory
StartupLoads the model into memoryReady when macOS is
Languages99, one model for allA shorter list, one system model per language
Model updatesWhen the app ships a new modelWith macOS updates
Choice of modelYes, in most appsNo
Intel MacsOften, slowlyNo: Apple Silicon only
Punctuation, fillers, correctionsNot the engine's jobNot the engine's job
The Whisper column describes whisper.cpp and WhisperKit based apps in general; details vary by app. Prices and features as checked on September 2026. Check the vendor's site before deciding.

Where the real difference lives

Both engines hand back words. Neither removes "um", applies "sorry, Thursday", puts the opening ¿ in a Spanish question, spells your colleague's name, or expands a snippet. That is the layer above, and it is where dictation apps actually differ from each other. A Whisper app with a good cleanup layer beats a SpeechAnalyzer app without one, and the reverse. What that layer adds, concretely.

The engine decides three things you feel every day: whether you download anything, how much memory stays taken while you work, and how many languages you get. It does not decide whether the text comes out ready.

What DeftNib chose and why

DeftNib is built on SpeechAnalyzer and puts its work into the layer above: 15 rule engines for the part that can be proven (a question is a question, a number is a number, a dictated address is an address), and Apple's local language model for the part rules cannot prove, with the rules checking its output. The trade is deliberate: no choice of model and a shorter language list, in exchange for an app of 9.7 MB that uses under 30 mb of RAM at rest, needs nothing downloaded, and improves when Apple improves the engine. The four reasons it is built this way.

Mac with Apple Silicon and macOS 26 or later. Free up to 4,000 words a week, no account.

By downloading you agree to the Terms and the Privacy Policy.

Questions

Is Whisper more accurate than Apple's engine?

It depends on the model size, the language and the audio, and there is no published comparison with a shared protocol that this page could cite. Both are good on clean English speech. The visible difference for a dictation user is usually the layer above the engine, not the engine.

Can I use Whisper on a Mac without downloading anything?

No. Whisper is a model, and a model is a file. Every Whisper app downloads at least one, from tens of megabytes to a few gigabytes depending on the size you pick.

Which one does DeftNib use?

SpeechAnalyzer, the engine Apple ships with macOS 26. There is no model inside the app and nothing to download after install; the app itself is 9.7 MB on disk.