Dictation accuracy: what affects it, how to improve

Dictation accuracy is affected by background noise, how clearly you speak, and whether the words are common or unusual. Parlo transcribes with Apple's on-device speech engine, then cleans filler sounds and false starts with an on-device language model. A custom dictionary protects typed words from autocorrect; it does not affect dictation. A few speaking habits matter more than any setting.

By Eric Friedman. Last updated 2026-09-06.

What actually affects how accurate a dictation is?

Four things matter most: background noise, how clearly and steadily you speak, whether the words you are saying are common or unusual, and whether the app has any way to learn the unusual ones. None of this is unique to Parlo. Every dictation system, on-device or cloud, deals with the same inputs.

What runs the transcription itself?

Transcription uses Apple's SpeechAnalyzer and SpeechTranscriber on the device. It is the only engine, with no cloud engine and no option to bring your own key. Since it runs locally, accuracy depends partly on Apple's model itself and partly on the factors above, rather than on network conditions or server load.

Does cleanup help or hurt accuracy?

Cleanup runs on the device with an on-device language model. It removes filler sounds, false starts, and repeated words, which usually makes a transcript read more like what you meant to say. It also applies spoken formatting commands. "New line," "new paragraph," and "parentheses" and "end parentheses" run through a deterministic pass that works even when the model itself is unavailable. "Quote" and "end quote" need the language model, so they only apply when AI cleanup is on. A length guard checks the cleaned-up result against what you actually said. If cleanup output is implausibly longer than the source, Parlo falls back to the deterministic text instead of guessing.

What does the custom dictionary actually do for dictation?

Less than the name suggests. The custom dictionary is edited under Settings, with an "Add word" option, and the terms you add sync as benign data across your devices. It is read only by the Parlo keyboard's autocorrect, as a do-not-correct list for words you type: add a name once, and the keyboard stops changing it back to a different word when you type it later. Dictation recognition itself does not read the dictionary. Apple's speech engine either recognizes a name from the audio or it does not, regardless of what you have added under Settings.

Why might accuracy feel different in different apps?

The keyboard inserts text after you dictate, but the transcription and cleanup happen in the app regardless of which app you swipe back to. If accuracy feels different across apps, the more likely cause is the environment you were in when you dictated, or a term specific to that context that is not yet in your dictionary, rather than the target app itself.

Does turning cleanup off make the raw transcript more accurate?

Not necessarily. "Use AI cleanup" is a toggle on both platforms, on by default on iPhone and off by default on Mac. Either way, turning it off skips the language model pass, but the deterministic parts of cleanup still run: filler sounds still get stripped, and "new line," "new paragraph," and "parentheses ... end parentheses" still apply. What you lose without AI cleanup is the model's pass at smoothing false starts and repeated words, plus the "quote ... end quote" command, which needs the model. Whether that makes a transcript feel more or less accurate depends on how you tend to speak. Turning the toggle off is a quick way to compare the two for yourself on a real dictation.

Does what you say out loud change how accurately it gets captured?

Yes, more than most settings do. A sentence spoken at a steady pace, finished before you pause, gives the speech engine a clean signal to work with. A sentence that trails off, restarts halfway, or gets interrupted mid-word is harder for any speech engine to resolve cleanly, on-device or in the cloud. This is not about speaking unnaturally slowly. It is about finishing a thought before the next pause, the same habit that helps a person understand you on a phone call with a bad connection.

Does a first-run download affect accuracy?

Indirectly. The first run may need a one-time download of Apple's on-device speech model. Parlo shows a message asking you to connect to the internet and try again if the model is missing. Until that download finishes, dictation cannot run at all, which can look like an accuracy problem the first time you try Parlo on a new device, when it is actually a one-time setup step.

Does the length guard ever hide a real mistake?

The length guard is a narrow safety check. It compares the length of cleanup's output against what you said. It falls back to the deterministic text when the output is implausibly longer, which usually means the model invented content rather than cleaned up what you actually said. It does not check every kind of error, so it is not a general accuracy guarantee. It catches one specific failure mode and leaves the rest to you to review.

How can you improve dictation accuracy in practice?

1. Add names, acronyms, and recurring terms to your custom dictionary once, so the keyboard's autocorrect stops changing them back when you type them.
2. Speak in a quieter space when accuracy matters, such as a call recap or a message with names in it.
3. Finish sentences before pausing, rather than trailing off mid-thought.
4. Use spoken formatting commands instead of trying to describe punctuation in words.
5. Turn "Use AI cleanup" off temporarily if you want to compare the plain transcript against the cleaned-up version for a specific dictation.

Does English-only support affect accuracy for non-English speakers?

Yes. Parlo is English-first, and transcription falls back to en-US when a requested locale is unavailable. If you regularly dictate in another language, accuracy will reflect that fallback rather than a dedicated model for your language. A language picker is not built.

The fastest way to judge accuracy for your own voice and vocabulary is to try it. The beta page has the install links for iPhone and Mac, both free during the beta.