How it's tested
How Precis is tested
Dictation apps promise accuracy and rarely show how they check it. Here is what I test in Precis, how, and what I haven't tested.
Precis is checked three ways. A fact guard throws away any rewrite that adds a name, number, date or link you didn't say. A speech test scores each engine on real recordings of 20 read phrases. A layout test runs real prompts through Apple's on-device model. All results are mine, from my own Macs.
The fact guard: rewriting without inventing
Apple's on-device model, asked to write freely, signs emails "John Doe" and answers dictated questions instead of rewriting them. I measured this while building Styles. So every rewrite passes through a guard before it can be typed. The guard throws the result away if it:
- answers what you said instead of writing it;
- adds a number you didn't say, or leaves out one you did;
- adds a day or a month you didn't say;
- adds a length of time or a time of day you didn't say ("for a few days");
- adds a name or other proper noun you didn't say;
- adds an email address or a link you didn't say;
- drops too much of what you said, or grows far longer than it.
When the guard rejects a result, or the model runs past its time limit (a few seconds, longer for longer dictations), Precis types the plain clean-up instead: ums removed, punctuation fixed, your words kept. Automated tests cover each of these rejections. In my own scorecard of 58 test dictations, run on 1 October 2026, Styles added no invented facts: 55 were rewritten for the app, and 3 fell back to plain clean-up (two in languages Apple's model doesn't support, one that timed out).
Speech engines: scored on real recordings
I record 20 phrases in my own voice and run every engine over the same recordings. The phrases are chosen to be hard: names (Nikesh, Priya, Karthik, Rahul), times and prices ("5:30 pm on Friday", "29 dollars"), self-corrections ("Tuesday, no wait, Wednesday"), fillers ("um, so basically I think we should, uh, launch it"), and tech terms (Supabase, Vercel, SwiftUI, GitHub). Each recording is matched to the phrase it is closest to, scored against what I read, and timed.
- NVIDIA Parakeet (English and 24 European languages): 0.08 to 0.17 seconds per sentence on my M4 Mac.
- OpenAI Whisper (99 languages): 0.67 to 1.84 seconds per sentence on the same Mac.
The default, Automatic, picks the engine that knows your language.
Silence: no speech, no transcription
Given silence, Whisper doesn't return nothing; it invents "Thank you." So Precis first compares your voice with the room's own noise, not a fixed loudness, which means a whisper still counts. If it finds no speech, nothing is transcribed and the recording is deleted. If your microphone is muted or sending nothing, Precis says so instead of a vague "No speech".
Styles: which layout does the model pick?
I run real prompts through the real on-device model: a short question, a rambling bug report, a request that lists constraints, a trip plan. The report records which layout the model chose (a paragraph, the ask first with bullets, or sections) and how many seconds it took, and I read every result by eye.
What I haven't tested
- One speaker. The recordings are my voice, on my microphones. Your accent, microphone and room will give different results.
- Speed on one Mac. The timings above are from an M4. Other Apple silicon Macs will differ.
- No head-to-head accuracy contest. I haven't scored Precis against Wispr Flow, Superwhisper or others on the same recordings, so I don't claim it's more accurate than they are.
- No outside audit. These are my own checks, not an independent test.
If Precis mishears you, use Send Feedback in the app, or fix the word once in the Dictionary. The quickest way to judge accuracy for your voice is the 7-day free trial.
Questions
Does Precis make things up when it rewrites?
It is built not to. Every rewrite passes a fact guard that rejects any name, number, date, time, email address or link you didn't say, and falls back to plain clean-up if it fails. In my own scorecard of 58 test dictations (1 October 2026), Styles added no invented facts; 3 of the 58 fell back to plain clean-up. That is my own measurement, not an independent audit.
How accurate is Precis?
I don't publish a single accuracy percentage, because it depends on your voice, microphone and room. I score each speech engine on real recordings of hard phrases, and the 7-day free trial is the fastest way to test it on your own voice.
How fast is dictation in Precis?
Speech recognition takes 0.08 to 0.17 seconds per sentence with Parakeet and 0.67 to 1.84 seconds with Whisper, measured on an M4 Mac. Rewriting for a Style adds a few seconds at most, and the first dictation after launch can be slower while the models load.
Has Precis been compared with Wispr Flow or Superwhisper?
Not on accuracy. I compare features, price and privacy on the comparison pages, and I re-read each competitor's own site before dating a fact. I haven't run a head-to-head speech accuracy test against them.
Try Precis free for 7 days
Hold ⌥ Space in any app and talk. No card and no account, then $19 once for 2 Macs.
Free for 7 days, no card. Needs macOS 26.5 on Apple silicon.