What macOS 26's SpeechAnalyzer changes for dictation on a Mac

What SpeechAnalyzer is

SpeechAnalyzer is the speech recognition engine Apple introduced with macOS 26: it transcribes audio on the device, on the Neural Engine, using language models that macOS downloads and updates by itself. It is exposed to developers through the Speech framework, and it is the engine behind system dictation, Notes, Voice Memos and Journal.

What it replaced

Before it, apps used SFSpeechRecognizer, an API from 2016 designed around short requests, with a time limit per recognition and, for many languages, a server round trip. It worked, and it shaped a generation of dictation features that felt like a mode you switched into rather than a key you held. SpeechAnalyzer is built for long audio, streams results as you talk, and does all of it on the Mac.

What it means for users

• Nothing to download from the app. The language model lives in macOS. An app built on SpeechAnalyzer can be a few megabytes, because the model is not inside it.
• Model updates come with macOS. When Apple improves recognition for your language, every app built on the engine improves the same day, without shipping anything.
• One engine, many apps. Notes, Voice Memos, system dictation and third-party apps hear the same way. What differs is what they do with the words.
• On the device, by design. Audio is processed on the Mac. An app that uses this engine does not need a server to transcribe, and a user can check that with Wi-Fi off. On-device vs cloud, in general.

What it does not do

It writes the words. It does not decide that the sentence was a question you did not label, drop the "um", apply "sorry, Thursday", put the opening ¿ in Spanish, spell your colleague's surname, or expand a snippet. That layer is the app's job, and it is where dictation apps built on the same engine end up very different from each other. Where the built-in dictation stops.

What DeftNib builds on top

DeftNib uses SpeechAnalyzer as its only speech engine and puts its work in the layer above: 15 rule engines for the part that can be proven (punctuation, numbers, addresses, lists, spoken corrections, the personal dictionary, snippets), and Apple's local language model for the part rules cannot prove, with the rules checking its output. Because the engine and the model are Apple's and already on the Mac, the app is 9.7 MB on disk, has 1 third-party dependency, and measures 88 ms at the 95th percentile for short dictations, from key release to text. Why it is built this way.

Requirements

macOS 26 or later on Apple Silicon. Apple Intelligence on for the model-based cleanup; without it, the rule engines still run. How it compares with Whisper-based apps.

Questions