The evidence
The tests.
And their limits.
Your own sentences are the test that matters. These measurements explain what has been checked so far, and what they cannot tell you.
From releasing the key to a transcript
0.618 s median
0.629 s maximum across six fixture replays on an M4 Max. Measured with the shipping default configuration on September 5, 2026.
This measures key release to transcript using replayed audio. It does not establish live microphone latency or the time until text appears in another app. The replay harness suppresses insertion. Six replays on one Mac are not a guarantee about your Mac or your voice.
The homepage animation is an illustration, not a recording of this test and not a timing claim.
German recognition
12.71% word error rate
144 German clips, three synthetic voices, the default configuration. All attempts were counted. All 144 produced text; that does not mean they were error-free or delivered into another app.
The audio is synthesized speech, not a study of different people in different rooms. The test includes a vocabulary correction pass. Most terms in that vocabulary also appear in the reference sentences, which makes the absolute result optimistic on this set. Word error rate also does not tell you whether a mistake changed the meaning of a sentence.
The audio, written references, and scorer are published in eval/de. The full measurement report includes the methods, configuration, and engine comparison.
Use by the person who built it
174 dictations across 26 days
The developer’s own use on an M1 Max recorded zero observed real failures across those dictations. This is one person’s use, not a customer count, an independent review, or a claim of perfect word recognition.
It is a reason to test the app yourself. It does not establish how PressTalk performs for every speaker or workflow.
Check the claims yourself
Read the measurement record
Get the German test set and scorer
Read the full privacy explanation
Inspect the source