Why Mac dictation gets your words wrong
Written 15 August 2026.
Most dictation accuracy complaints are not one problem. They are three, they have different causes, and only one of them is about how you speak.
The short version: if it mangles names and jargon, the model has never seen those words and you need to give them to it. If it mangles whole clauses, it is probably the wrong language model for what you actually speak. If the words are right but the text is unusable, that is punctuation and formatting, which is a separate stage entirely.
1. It has never heard the word
A speech model transcribes by fitting sounds to words it knows. Hand it a colleague's surname, an internal product name, a drug name, a ticker or a library, and it will pick the nearest thing in its vocabulary. That is not a malfunction — it is the model doing exactly what it does, with a word list that does not contain your word.
The fix is to add the word, and the reason it works is worth understanding: a custom dictionary term is fed to the recogniser before it transcribes, biasing it toward your spelling. It is not a find-and-replace after the fact. That is why adding five names you say every day tends to fix more than any amount of speaking more clearly.
It also means the order matters. In Bunso, dictionary terms bias the recogniser, corrections are applied to the transcript after it, and saved keywords expand after cleanup. An earlier stage always shadows a later one, which is why the app validates all three lists against each other when you add anything — a keyword whose words a correction already claims can never fire.
2. It is the wrong model for the language you actually speak
This is the failure people misread as a bad accent. If you code-switch — English and Filipino, or Hindi, or Spanish, mixed inside a single sentence — the recogniser is being asked to do something a language setting cannot express. You pick one language before you speak; the mixing happens mid-sentence.
The result is not a slightly worse transcript. On a thirty-clip test of everyday Taglish, macOS dictation lost the Filipino content entirely in 28 of 30 clips, replacing it with English-shaped words that sound similar and mean nothing. That is structural: the current macOS dictation engine ships thirty locales and none of them is Filipino.
The full thirty transcripts are published, both systems, including the clip where macOS heard a word we missed. If your speech is entirely English, none of this applies to you and the built-in dictation is a perfectly good tool.
3. The words are right and the text is still unusable
Speech recognition gives you what you said. What you said includes um, sorry, let me start again, the sentence you abandoned halfway, and no punctuation. Every one of those is a faithful transcription and none of it is text you would send.
Fixing that is a second pass over the transcript, separate from recognition. It is also the stage that can do damage: a pass that rewrites too eagerly will "correct" a technical term into a common word, or answer a question it thinks you asked. What that pass should and should not touch is its own subject.
Which one do you have?
A quick diagnostic. Dictate the same three sentences: one with only common English words, one loaded with the names and jargon you use at work, and one the way you would actually say it to a colleague.
- All three fine except the jargon one — vocabulary problem. A custom dictionary fixes it, in any tool that has one.
- The natural one falls apart — language problem. You need a model that covers what you actually speak, not a better microphone.
- All three accurate but nothing is punctuated — formatting problem. You need a cleanup pass, not better recognition.
The reason to separate them is that the usual advice — speak slower, enunciate, buy a better microphone — only ever helps with a fourth problem most people do not have.