Measured on public tests, in public. Here is what they say, and what they don't.
Dictation apps rarely publish accuracy. We do, because “it runs on your Mac” only matters if it is also good. The numbers below are from version 1.8.16. The benchmark scripts and the exact clip lists exist and can be re-run; if you want to check a figure, ask and I will share them.
Dictation
The Open ASR Leaderboard is the standard public comparison of speech-to-text models. We ran 300 clips from each of its test sets — meetings, earnings calls, podcasts, audiobooks, parliamentary speech — through the whole app, exactly as it works on your Mac, and scored them the leaderboard's way (word error rate, lower is better).
| Test set | Mythral, engine | NVIDIA Parakeet, published |
|---|---|---|
| AMI (meetings, headset mics) | 7.8% | 11.2% |
| Earnings-22 (earnings calls) | 12.4% | 11.2% |
| GigaSpeech (podcasts, audiobooks) | 9.9% | 9.7% |
| LibriSpeech clean | 1.8% | 1.7% |
| LibriSpeech other | 3.6% | 3.2% |
| SPGISpeech (financial) | 2.1% | 2.2% |
| VoxPopuli (parliament) | 7.2% | 6.0% |
| Average | 6.4% | 6.4% |
In plain terms: the app hears as well as the model it runs, which is one of the best open English models available, and it does so on your laptop. After the app's own clean-up (stutters folded, “you know” removed — the things you would want out of dictated text) the figure is 7.0%, because the test transcripts are verbatim and count that clean-up as error. The chart on the home page uses the leaderboard's newer results file (19 September 2026), where the same model averages 5.9% over the full test sets; the 6.4% here is its published figure at the time of this test, next to ours on the same 300 clips.
Meetings
Sixteen four-person meetings from the AMI corpus, 147 minutes in all, run through the same path the app takes when a meeting ends on your Mac:
| Words wrong | 10.5% (Parakeet's own figure for this material is 11.2%) |
|---|---|
| Words kept | 95% — the rest is repeated back-channel (“yeah yeah yeah”) folded on purpose; no stretch of speech goes missing |
| Right speaker on a line | 92% of what was said |
| Time to write up a 6-minute meeting | About 7 seconds |
What the numbers leave out
- These are samples. 300 clips per set, not the full sets, so each figure carries roughly ±0.5 points of uncertainty. They are not leaderboard submissions.
- Very short replies are the weak spot. A quarter-second “yeah” or “okay” is still wrong about half the time. Longer than a second, it is fine.
- Names it has not been taught. On earnings calls, given the speakers' own names in advance, it spells about two in three exactly right. The dictionary, Contacts import and “add word” exist for the third.
- Accents. The public sets are largely American and British English. Other accents will do worse; we have not measured by how much.
- Speaker counting. In a room of four it sometimes hears five or six voices; the label on a line is still right 92% of the time.
What the measuring found
Running the benchmark found real bugs, which is the point of it: speech from quiet or distant microphones was being discarded as silence, one-word replies came back empty, and the word “period” was becoming a full stop in the middle of financial sentences. All fixed in 1.8.16. The benchmark runs again before every accuracy-related release.