"How accurate is it?" is the first question anyone asks about a dictation tool, and it's the wrong shape of question. Accuracy sounds like a single number — a percentage you could put on a box — but no one number tells you whether a tool will actually save you time. A tool can catch nearly every word you said and still hand back text you'd never send. It can transcribe a slow, clean sentence in a silent room and fall apart on the way you actually talk. And the "supports 100+ languages" line tells you nothing about how it does on yours. What you need isn't someone else's benchmark — it's a way to judge a tool on your own speech, in a few minutes, and to read the result honestly. Here's the framework.
Word accuracy is the floor, not the ceiling
The number people mean by "accuracy" is usually word accuracy: of all the words you said, how many did the tool get right. It matters — a tool that mishears the words has nothing to build on — but it's the floor of the question, not the whole of it. Word accuracy measures whether the sounds were caught, not whether what came back is text you can use.
Two things make it less informative than it looks. First, it degrades exactly where you need it most: on names, on jargon, on the acronym your team uses ten times a day, on the product name that isn't in any dictionary. A tool can score beautifully on ordinary prose and still miss every proper noun in your sentence — and those are the words a reader notices. Second, not every wrong word costs the same. In some languages a mistaken character is a different word entirely, not a typo you skim past. So the same accuracy figure can describe a result you'd happily send or one you'd have to comb through. The percentage doesn't distinguish them. Your own eyes will.
Formatted output is a different measurement
Speech doesn't arrive with punctuation, paragraphs, or capital letters. You don't say "comma" or "new paragraph" — you pause, you change tone, you move to the next thought. Turning that stream into readable text is a second job, sitting on top of catching the words, and it's where the gap between tools is widest.
This is the difference between a transcript and something you'd send. A wall of words with every clause caught correctly but no sentence breaks is technically accurate and practically useless — you still have to go back and punctuate it. Formatted-output accuracy asks a harder question: did the sentences break where you meant, did the right homophone get chosen from context, did a spoken list come back as a list? When you test one, judge the shape of the text as strictly as the words.
"Kept my meaning" is the one that decides it
The highest bar, and the one that actually decides whether you'll keep using a tool, is whether the finished text says what you meant. That isn't the same as getting the words right. You can have every word transcribed correctly and still get back something clumsy or ambiguous — because spoken language and written language are different registers. We repeat ourselves, start over mid-thought, lean on "um" and "like," and trail off. Read verbatim, natural speech looks worse on the page than you'd expect.
A tool that only transcribes hands you all of that, faithfully. A tool that rewrites toward clean text has to walk a line: tidy the speech into something readable without changing what you said. Clean up too little and you're left editing; clean up too much and the tool "improves" your sentence into one you didn't mean — softens a firm no, guesses a fact you never stated, flattens a deliberate word choice. When you test, watch this hardest: read the output against your memory of what you said and ask not "are the words right" but "is this still mine." A tool that keeps your meaning while making you read better is doing the whole job; one that changes your meaning to make you read better is doing something you can't afford.
A language count is not a quality claim
"Supports 100+ languages" is a coverage claim, and coverage is the easiest thing to list and the hardest to verify. A language on the feature page means the tool will attempt it. It says nothing about whether the attempt comes back usable — and quality varies enormously across the languages a single tool claims. A tool can be genuinely excellent in the two or three languages it was built around, and thin the moment you move down the list to yours.
The languages most likely to be thin are the ones that were hard to build well: large homophone spaces that need sentence-level context, writing systems with no spaces between words, speech mixed with English in every other clause, more than one written form for the same sounds. None of that shows up in a count. The only thing that tells you how a tool does in your language is running your language through it — so if your work happens in a language that isn't English, treat the feature-page number as marketing and the test below as the truth.
Test it on your own speech in five minutes
You don't need a benchmark or anyone else's numbers. You need one honest session. Pick something you'd genuinely write — a message to a colleague, a paragraph of a doc, a reply you owe someone — and dictate it the way you actually talk: your real pace and accent, the way you mix languages if you do, the names and jargon your work involves, the false starts you'd never rehearse away. Don't perform a clean sentence for the tool. The clean sentence isn't the test; your normal speech is.
Then read what comes back, against three questions in order. Are the words right, especially the names and terms a reader would notice? Is it formatted — punctuated, broken into sentences, shaped like writing rather than a transcript? And does it still say what you meant, without softening or inventing anything? A tool that clears all three reads like something you'd send without opening the editor. A tool that clears one or two has moved your work rather than removed it — you've traded typing for correcting, or dictating for checking that your meaning survived. That reading, on your own speech, tells you more than any accuracy percentage could. It's also, not by coincidence, the bar we hold Sageio Type to: not the highest score on someone's benchmark, but text you'd send as it came back. Run the test on whatever you're considering, and judge it the same way.