Fact · Accessibility and health
3.80percentage points of word error rate
In our October 2026 test, Whisper large-v3-turbo's word error rate on Italian readings of a short passage was 3.80 percentage points higher for healthy speakers aged 60 to 77 than for healthy speakers aged 19 to 29 (a measured difference: its interval, allowing for the test's comparisons, excludes zero).
Measured by us on October 23, 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
- OpenAI Whisper large-v3-turbo, 5-bit quantized (ggml-large-v3-turbo-q5_0.bin), run through whisper.cpp 1.9.4 with Metal on the GPU, with the language given. Both groups read the same printed Italian text in the same room setup (Italian Parkinson's Voice and Speech, University of Bari and ASL Bari).
- How it is measured
- Difference of the two groups' pooled word error rates (+3.80 points; 2.25 times the rate). Interval: 5,000 bootstrap draws of persons in each group separately (seed 7); 95% for this comparison alone +1.82 to +6.11, and allowing for all 9 comparisons (Bonferroni) +1.21 to +7.52 points, the interval the verdict stands on.
- What it does not show
- Groups of 15 to 28 speakers from one region of Italy reading a fixed text. A difference also includes reading slips (a skipped or changed word counts as an error), which may be more common in one group.