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

Latvian · 8 models · left is slower, down is better · log speed axis

AMD Ryzen 4750U GPU (Vulkan)
pareto frontier (no model is both faster and more accurate) other models
1 canary-1b-v2!overall pickmost accuratefast pick 10.419.79–11.09 13.2× 6.7× 1.10 GB CC-BY-4.0
2 whisper-large-v3 19.2118.35–20.07 2.1× 0.6× 1.55 GB Apache-2.0
3 whisper-large-v3-turbo 19.5318.64–20.43 3.4× 0.8× 0.83 GB Apache-2.0
4 parakeet-tdt-0.6b-v3 23.7722.87–24.77 12.5× 7.5× 0.72 GB CC-BY-4.0
5 parakeet-primeline 28.8327.83–29.95 12.5× 7.5× 0.72 GB CC-BY-4.0
6 whisper-medium 33.5832.56–34.62 4.3× 1.1× 0.77 GB Apache-2.0
7 Fun-ASR-MLT-Nano-2512! 56.7155.13–58.29 9.0× 4.5× 0.83 GB FunASR Model Open Source License Agreement v1.1
8 whisper-small 57.4256.20–58.69 12.1× 3.4× 0.25 GB Apache-2.0