Is speech recognition less accurate on long recordings?

Sometimes, and it depends on the engine. We gave three engines the same 84 English readings of 3 to 4 minutes, once as a single recording and once cut at pauses into pieces of 10 to 20 seconds. Whisper large-v3-turbo got worse on the long recording: its word error rate went from 2.35 percent word error rate ¹ in pieces to 2.91 percent word error rate ² as one recording. Parakeet TDT v3 got slightly worse. Apple SpeechTranscriber did not change in any way this test can detect.

Two of three engines made more errors on the long recording

The test measures one thing: what handing over a longer recording costs each engine, with the audio and the reference text held the same. The cost is the word error rate of the one recording minus the word error rate of the pieces.

• Whisper: 0.56 percentage points of word error rate more errors on one recording ³. This is a measured difference.
• Parakeet: 0.25 percentage points of word error rate more ⁴. This is also a measured difference, but a smaller one, and its margin of error only just clears zero.
• Apple: 0.01 percentage points of word error rate more ⁵. The margin of error spans zero in both directions, so we treat Apple as unchanged. We do not claim it is better or worse.

The margins of error are intervals that allow for the test making several comparisons at once. Whisper's sits well above zero. Parakeet's starts barely above it. Apple's runs from below zero to above it. Each row's note on the full dataset page gives the exact range.

Whisper also lost whole passages on long recordings

Averages hide a more disruptive failure: an engine skipping a passage. We counted readings where an engine left out 10 or more words in a row.

Whisper lost a passage only when it got the long recording, never in pieces. Apple lost none in either condition. Parakeet lost passages in a few readings under both conditions, and with counts this small we do not read a pattern into which condition had more.

On short pieces, the engines are level

Cut into pieces, the three engines land close together: Apple at 2.33 percent word error rate ¹², Whisper at 2.35 percent word error rate ¹ and Parakeet at 2.45 percent word error rate ¹³. None of the pairwise gaps is a measured difference, so the test counts them as level (for example Apple against Parakeet ¹⁴).

On one recording the picture changes. Apple's word error rate was lower than both Parakeet's ¹⁵ and Whisper's ¹⁶, and both gaps are measured. Parakeet against Whisper stays level ¹⁷. The gap between engines on long audio is therefore mostly a length effect, not a ranking that holds for short audio too.

What these numbers do and do not measure

The readings are audiobook recordings from LibriSpeech-Long: clean read English, 40 speakers, one reading per chapter. They are easier than dictating your own words at a desk, with hesitations, corrections and a background. Absolute error rates on your speech will differ. The test does not cover other languages, noisy audio (see our noise robustness data), streaming while you speak, or cloud engines.

It also isolates length. An app that cuts or streams your audio itself may behave like the "pieces" condition even when you speak for minutes. For how the error rate itself is defined, read our guide to speech recognition accuracy.

What it means for choosing how to dictate

• If you transcribe long files with Whisper, splitting the audio at pauses is worth doing. Its error rate and its lost passages both improved in pieces.
• With Apple's engine, length made no detectable difference here, so you can hand over a long recording as is.
• Whatever you use, skim long transcripts for a missing sentence, because a skipped passage does not show up in an error-rate average.

See how to transcribe a long audio file and how to dictate a book or long document on iPhone for practical steps. The full dataset lists every reading condition. For the wider subject, start with our overview of speech recognition, and see the on-device benchmark for engines on other tests.