Can you dictate for several minutes without stopping?

For the recognition itself, yes. We gave three engines 84 English readings of 3 to 4 minutes, each as one continuous recording, and all three transcribed them with low error rates: Apple SpeechTranscriber at 2.34 percent word error rate ¹, Parakeet TDT v3 at 2.70 percent word error rate ² and Whisper large-v3-turbo at 2.91 percent word error rate ³. Whether your dictation app lets you speak that long is a separate question that this test does not answer.

Continuous speech was transcribed accurately by all three engines

Lower is better, and the three figures above are all small. On clean read speech, only a small share of words went wrong. The test used one recording per reading, with no pauses cut out and no pieces rejoined.

Apple's word error rate on one recording was lower than Parakeet's ⁴ and lower than Whisper's ⁵, and both gaps are measured differences. Parakeet and Whisper count as level with each other ⁶.

Whisper skipped passages on long recordings, the others rarely or never

An error rate averages over the whole reading. A skipped passage is worse for a dictating person, because the missing sentence leaves no visible mistake. We counted readings where the engine dropped 10 or more words in a row.

Apple lost none. Whisper lost a passage in 4 of 84 readings as one recording and in none when the same audio was cut into pieces ¹⁰.

Length cost something to Whisper and Parakeet, nothing detectable to Apple

Against the same audio in short pieces, one recording raised Whisper's word error rate by 0.56 percentage points of word error rate ¹¹ and Parakeet's by 0.25 percentage points of word error rate ¹². Both are measured differences. Apple's change was 0.01 percentage points of word error rate ¹³, which the test cannot tell from chance. So for a few minutes of speech, the engine decides how much you lose to length.

What this test does not cover

• An app's own limits. Many dictation tools stop after a pause or after a set time, or cut your audio before recognition. We tested engines on finished recordings, not any app's recording behavior.
• Streaming. Dictating live, where words appear as you speak, is a different mode from transcribing a finished recording. We did not measure it.
• Your own words. The readings are audiobook text read aloud in a quiet room. Composed speech has false starts and unusual vocabulary, so expect higher error rates. The test also covers only English and no cloud engines.
• Long sessions. The readings stop at about 4 minutes, so the test says nothing about a much longer session without a break.

Our noise robustness data shows how quickly accuracy changes once a room is not quiet.

Practical advice for long dictation

1. Dictate in natural paragraphs and pause between them. Pauses are where an app or a file splitter can cut safely.
2. If your tool uses Whisper, prefer one that splits at pauses before recognition, since pieces avoided lost passages in our test.
3. Read the result once for missing sentences, not only for misspelled words.
4. Check your app's settings or documentation for a maximum recording time before you start a long session.

Yap, which we make, runs dictation on the device, but this page reports engines, not our app. For tools compared side by side, see the market overview. For step-by-step guidance, see how to dictate a book or long document on iPhone and how to transcribe a long audio file. The full dataset shows every row, and our guides to speech recognition and speech recognition accuracy explain the measure. For other tests of the same engines, see the on-device benchmark.