Fact · Speech recognition
15.52percent formatted error rate
In our October 2026 benchmark, counting punctuation and capital letters as a reader sees them, Parakeet TDT 0.6B v3's text differed from the written Croatian sentences by 15.52 errors per 100 tokens, against 12.44 for the words alone.
Measured by us on October 19, 2026, from 100 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 reference is FLEURS's written sentence, with its punctuation and capitals. The gap to the plain word error rate is the editing a dictated text needs for punctuation and capitals alone.
- How it is measured
- Formatted error rate = 332 errors / 2,139 reference tokens = 15.52 percent, over 100 recordings; tokens are words with their capitals plus each sentence mark (. , ! ? ; :) on its own.
- What it does not show
- FLEURS's sentences follow their source's punctuation, and a comma can be optional, so part of the gap is style rather than error, for every engine alike. Numbers written as digits against words count as errors.
In the data
A row of Dictation Punctuation and Capitalization Accuracy, as Parakeet TDT v3, Croatian: 15.52 percent.
- Engine
- Parakeet TDT v3
- Maker
- NVIDIA
- Language
- Croatian
- Words only (WER)
- 12.44
Used to calculate
- In the formatted accuracy of on-device speech recognition: punctuation and capitals counted dataset, the 25 published figures for “Parakeet TDT v3” have a median of 12.75 percent, ranging from 5.44 to 34.83.
- The formatted accuracy of on-device speech recognition: punctuation and capitals counted dataset holds 56 data points, each a published figure traced to its source.