Most people picture dictation as a single act: you talk, and words appear. For a long time that was accurate, and it's exactly why voice typing had a reputation for producing text you'd never actually send — a runny stream of lowercase words, no commas, every "um" faithfully preserved, "new paragraph" typed out as three literal words. What changed recently isn't that the microphone got better. It's that a second stage was added after the listening. Modern AI dictation is two models doing two different jobs: one turns sound into a rough transcript, and one turns that rough transcript into text a human would write. Almost everything you notice about whether dictation feels usable comes down to that second stage — so it's worth understanding both.
Stage one: speech to a raw transcript
The first job is transcription, handled by a speech-recognition model. It takes the audio of your voice and produces a best-guess string of words. This is the part that has existed, in some form, for decades, and on a clean sentence spoken slowly in a quiet room it's remarkably good. But its output is raw by design: it's a record of the sounds it heard, in order, with little sense of where a sentence should end or which of several identically-sounding words you meant. It doesn't know that the pause you took was the end of a thought, or that "period" was an instruction rather than a word. It hears sound and emits a transcript. On its own, that transcript is the "court reporter" text — technically a correct log of what was said, and almost never what you'd want to send.
Stage two: transcript to finished text
The second job is the one that makes the difference, and it's newer: a language model reads the raw transcript and rewrites it into finished prose. This is where the commas and capital letters come from — the whole business of punctuation and formatting being inferred rather than spoken aloud. It's where "um," "you know," and the false start you abandoned halfway get quietly dropped. It's where a rambling spoken sentence — the kind everyone produces, because speaking and writing are different skills — gets restructured into something with a subject, a verb, and an ending. The transcript says what you said; this stage produces what you meant to write. A tool that skips it hands you stage-one output and calls it dictation. A tool built around it treats the transcript as raw material, not as the answer.
Why the rewrite pass is the whole game
The gap between those two philosophies is the gap between dictation that feels like 2024 and dictation that feels like now. Speaking is fast and loose: you backtrack, you insert filler while you think, you trail off and restart. Writing is tight and structured. Bridging that gap used to be your job — you dictated, then you went back and cleaned up, which often cost more than just typing would have. Moving that cleanup into the tool is what makes speaking a genuine substitute for typing rather than a first draft of it. When people say a dictation tool "just works," what they're describing, almost always, is a strong rewrite pass: they spoke messily, the way people do, and clean text came out.
Tone is part of the rewrite
The rewrite stage isn't only fixing errors — it's making choices, and the most important one is register. The same spoken sentence should not become the same written sentence everywhere. A thought you speak while a chat window is open should come out loose and short; the same thought aimed at a formal email should come out composed and complete. Good dictation notices where the text is going and matches the tone to it, because "make it sound right" means something different in a message to a colleague than in a note to a client. This is judgment, not transcription, and it's only possible because a language model — not just a speech recognizer — is doing the writing. It's also the clearest tell of a tool's generation: fixed, one-size output is a stage-one tool with a spell-checker; tone that shifts to fit the destination is stage two doing its actual work.
Where languages other than English raise the stakes
Both stages get harder outside English, and the rewrite stage gets more valuable. In languages with dense homophones, the transcription stage genuinely cannot pick the right written form from sound alone — it needs a model reasoning over the whole sentence's meaning to choose correctly. In languages with honorific or politeness levels, the "tone" the rewrite picks isn't a nicety, it's grammar that changes across the sentence. And in the mixed-language speech that's normal in a lot of workplaces, it's the second stage that has to reassemble one clean sentence from audio that switched languages mid-clause. So the two-stage design isn't an English luxury with extras bolted on for everyone else; the harder the language, the more the finished text depends on the model that writes rather than the one that merely hears.
What this means when you're choosing a tool
Once you can see the two stages, the questions worth asking get sharper, and they map onto the practical options on a Mac — built-in dictation, browser tools, dedicated apps. A feature list that says "voice typing" tells you there's a stage one; it tells you nothing about stage two, which is where usable text is won or lost. So don't test dictation by reading a clean sentence in a quiet room — that only exercises transcription. Test it the way you'll use it: speak the way you actually speak, with the backtracks and the filler and your normal mix of languages, aimed at the app you'd really be writing in. Then read what comes back and ask whether it's send-ready or draft-ready — the same discipline behind judging accuracy for yourself instead of trusting a benchmark. That single reading tells you which of the two tools you're holding — the one that logs your voice, or the one that writes for you.