Fact · Speech recognition
0.25percentage points of word error rate
In our October 2026 test, Parakeet TDT 0.6B v3's word error rate on English readings of 3 to 4 minutes was 0.25 percentage points higher when each reading was handed over as one recording than when the same audio was cut into pieces of 10 to 20 seconds (a measured difference: its interval, allowing for the test's comparisons, excludes 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
- NVIDIA Parakeet TDT 0.6B v3 converted to Core ML (Parakeet Ultra, FluidInference, model revision 95eaa59a39d4), run through FluidAudio 0.17.4 on the CPU and GPU, with the language given. This is the engine the Yap app uses on Mac. 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.
- How it is measured
- One recording's pooled word error rate minus the pieces' (+0.25 points), on the same 84 readings. Interval: 5,000 bootstrap draws of speakers, paired (seed 7); 95% for this comparison alone +0.06 to +0.45, and allowing for all 3 engines (Bonferroni) +0.02 to +0.49 points, the interval the verdict stands on.
- What it does not show
- The pieces were cut at pauses chosen in the audio (the quietest tenth of a second in each 10 to 20 second stretch); 2 of the 1,255 cuts fell where the reader barely paused (less than 15 dB below speech), which, if anything, costs the pieces condition words.