Fact · Accessibility and health
1.26percentage points of word error rate
In our October 2026 test on Italian readings of ten short sentences by healthy speakers aged 60 to 77, Parakeet TDT 0.6B v3's word error rate was 1.26 percentage points higher than Whisper large-v3-turbo's (not a measured difference: its interval, allowing for the test's engine comparisons, includes zero, so the two engines count as level here).
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
- Both engines transcribed the same recordings (Parakeet TDT v3: 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. Whisper large-v3-turbo: 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.) Whisper ran with greedy decoding and timestamps on, the way whisper.cpp moves through recordings longer than its 30-second window (most of these are 50 to 250 seconds). A first pass with timestamps off lost most of several recordings and is not used. whisper.cpp's beam search was not tested.
- How it is measured
- Difference of the two engines' pooled word error rates on the same recordings (+1.26 points). Interval: 5,000 paired bootstrap draws of persons (seed 7); 95% for this comparison alone +0.09 to +2.52, and allowing for all 15 engine comparisons (Bonferroni) -0.36 to +3.05 points, the interval the verdict stands on.
- What it does not show
- One language, read text and 15 to 28 speakers per group: an engine's lead here may not hold for other voices or for people dictating their own words.