Dictation that finally works: what changed, and what it is still bad at

Dictation failed for twenty years on one thing. Not the grammar, not the speed, not the accents: the proper nouns. A tool that transcribes your sentence perfectly and mangles the name of your product, your client and your colleague produces text you have to correct word by word, and correcting is slower than typing was. What changed is that the tool now learns your names, and that single fix turns a gadget into something used fifty times a day.

Where dictation wins outright

Long messages where the thinking is already done. A briefing, a detailed reply, a specification, a summary of a call: the parts where you know what you want to say and typing is pure transcription cost. Speaking those takes a third of the time, and the output tends to be more direct, because people write like bureaucrats and speak like humans.

It also wins on the messages you were not going to write at all. The long reply that deserves ten minutes and gets postponed for three days often gets said in ninety seconds, and the delay disappears.

Where it is still bad

Anything dense in symbols. Code, formulas, file paths and structured data are worse dictated than typed, and no amount of model improvement changes that: the mapping from speech to punctuation is genuinely ambiguous. Short messages too: for six words, the overhead of starting a dictation is longer than typing them.

And text that needs to be precise on the first pass, like a legal sentence or a price. Spoken language is looser than written language, and dictated text carries that looseness. It is a first draft at speaking speed, not a finished paragraph.

The dictionary is the whole product

Learning your names and jargon is what makes the output usable without correction, and a shared dictionary spreads that across a team so that nobody has to teach the tool the same product name twice. If you are evaluating tools in this category, test exactly this: dictate a paragraph containing five of your own proper nouns and count the mistakes. Everything else is table stakes.

What it costs

Free for an individual, with speech to text in any application across more than a hundred languages on Mac, Windows, iOS and Android. Pro is fifteen dollars per user per month, twelve billed annually, and buys unlimited dictations, the shared dictionary and team management. Growth is twenty-three, eighteen annually, and adds SSO and admin controls. Read 18 September 2026 on the United States price list: their site displays in the currency of the visitor, which is worth knowing if the number on your screen looks unfamiliar.

The free tier being genuinely usable is the unusual part. It means the question is not whether to try it but whether the shared dictionary and the team controls are worth paying for, and that answer arrives naturally once more than one person is using it.

Where it fits next to an assistant

Dictation and AI writing solve opposite halves of the same problem. An assistant produces text you did not write, which then has to be checked against what you meant. Dictation produces your own words at speaking speed, which needs no checking for intent, only for form. For anything carrying a commitment, a price or a relationship, the second is safer and usually faster.

The combination that works: speak the substance, then hand the rough text to a model for structure if the message deserves it. Starting from a model draft and correcting it back towards what you actually meant is the slow path, and it is the one most people default to.

The adjustment nobody warns you about

You have to learn to speak in finished sentences, and that takes about a week. The first attempts are full of restarts because writing has trained everyone to edit while composing, which you cannot do out loud. Once the habit settles, the speed is real, and the writing gets shorter, because saying a sentence that runs for four lines is physically uncomfortable in a way that typing one is not.