Fact · Speech recognition
-0.31percentage points of word error rate
In our October 2026 test, handing over a 3 to 4 minute English reading as one recording instead of in pieces raised Parakeet TDT 0.6B v3's word error rate by 0.31 percentage points less than Whisper large-v3-turbo's (not a measured difference: its interval, allowing for the test's comparisons, includes zero).
Measured by us on October 25, 2026, from recordings
Our own measurement
This number comes from our benchmark, not from another publication. The full record (the setup, every recording's checksum, every transcript and timing) is published, so you can recompute it.
- What it measures
- 84 examples of 3 to 4 minutes read by 40 speakers from public-domain audiobooks (LibriSpeech-Long test-clean, Google DeepMind, from LibriSpeech), clean recordings of read English. Each engine's own two conditions share their audio and reference, so this compares what length does to each engine, free of how each writes numbers.
- How it is measured
- (One recording minus pieces) for Parakeet TDT v3 minus the same for Whisper large-v3-turbo, on the same readings. Interval: 5,000 paired bootstrap draws of speakers (seed 7); 95% for this comparison alone -0.62 to -0.02, and allowing for all 3 such comparisons (Bonferroni) -0.69 to +0.04 points.
- What it does not show
- Clean read English; dictating your own words, other languages or a noisy room may change how much length costs.