Fact · Accessibility and health
5.30percent word error rate
In our October 2026 test on a Mac with an M2 Pro, Parakeet TDT 0.6B v3 had a word error rate of 5.30 percent on 21 Italian readings of ten short sentences by 21 healthy speakers aged 60 to 77.
Measured by us on October 23, 2026, from 21 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. The speakers read a printed text aloud in a quiet room with a microphone 15 to 25 cm away (Italian Parkinson's Voice and Speech, University of Bari and ASL Bari, recorded 2016 and 2017). The reference is the text they were given, so a word a speaker skipped or changed counts as an error for every engine alike.
- How it is measured
- Word error rate = 59 word errors / 1,113 reference words = 5.30 percent, over 21 recordings (10.4 minutes of audio). The 95% interval over speakers is 3.41 to 7.37 percent.
- What it does not show
- 21 speakers from one region of Italy, in one language, reading a fixed text: a person dictating their own words, in another language or with more advanced Parkinson's may fare differently. Half of these speakers had an error rate of 3.8 percent or less, the highest was 17.0 percent.
In the data
A row of Dictation Accuracy With Parkinson's Speech, Tested, as Parakeet TDT v3, Healthy speakers aged 60 to 77, read phrases: 3.41 to 7.37 percent.
- Engine
- Parakeet TDT v3
- Maker
- NVIDIA
- Speakers
- Healthy speakers aged 60 to 77
- What they read
- Read phrases
- Speakers in group
- 21
- Recordings
- 21
- Language
- Italian