handy models

speech-to-text benchmarks for transcribe.cpp, the engine behind Handy

best model by language

Benchmarked on the FLEURS test split at Q8_0. Error is WER — lower is better. Speed is × realtime on Handy target hardware.

accuracy vs speed

Ukrainian · 16 models

AMD Ryzen 4750U GPU (Vulkan)
pareto frontier (no model is both faster and more accurate) other models
highlighted rows are recommended picks or on pareto frontier
1 parakeet-ultraoverall pickaccurate pickfast pick 5.71 23.0× 13.3× 0.69 GB CC-BY-4.0
2 whisper-large-v3 6.28 2.0× 1.0× 1.55 GB Apache-2.0
3 orukeet 6.53 25.1× 13.6× 0.69 GB CC-BY-SA-4.0
4 parakeet-tdt-0.6b-v3 6.84 25.1× 13.7× 0.69 GB CC-BY-4.0
5 whisper-large-v3-turbo 7.31 3.1× 1.4× 0.83 GB Apache-2.0
6 parakeet-primeline 8.11 25.0× 13.6× 0.69 GB CC-BY-4.0
7 whisper-large-v2 8.17 2.0× 1.0× 1.55 GB Apache-2.0
8 whisper-large 9.87 2.0× 1.0× 1.55 GB Apache-2.0
9 canary-1b-v2! 10.58 14.1× 8.3× 1.07 GB CC-BY-4.0
10 whisper-medium 11.59 4.1× 1.9× 0.77 GB Apache-2.0
11 moonshine-base-uk 14.38 26.8× 17.4× 0.07 GB MIT
12 nemotron-3.5-asr-streaming-0.6b 14.88 17.1× 10.6× 0.70 GB OpenMDW-1.1
13 moonshine-tiny-uk 18.76 45.0× 30.7× 0.03 GB MIT
14 whisper-small 20.42 11.7× 5.9× 0.25 GB Apache-2.0
15 whisper-base 42.03 29.0× 19.3× 0.08 GB Apache-2.0
16 whisper-tiny 57.49 43.1× 39.8× 0.04 GB Apache-2.0