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

French · 13 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 Qwen3-ASR-1.7Bmost accurate 4.524.06–5.01 3.8× 2.0× 2.04 GB Apache-2.0
2 canary-1b-v2!overall pickfast pick 5.094.61–5.59 13.2× 6.7× 1.10 GB CC-BY-4.0
3 cohere-transcribe-03-2026! 5.234.72–5.77 8.0× 3.0× 2.41 GB Apache-2.0
4 parakeet-tdt-0.6b-v3 5.304.77–5.78 12.5× 7.5× 0.72 GB CC-BY-4.0
5 whisper-large-v3 5.394.88–5.94 2.1× 0.6× 1.55 GB Apache-2.0
6 whisper-large-v3-turbo 5.515.01–6.06 3.4× 0.8× 0.83 GB Apache-2.0
7 parakeet-primeline 6.355.84–6.89 12.5× 7.5× 0.72 GB CC-BY-4.0
8 Qwen3-ASR-0.6B 7.767.14–8.43 8.0× 4.3× 0.79 GB Apache-2.0
9 whisper-medium 8.077.43–8.80 4.3× 1.1× 0.77 GB Apache-2.0
10 canary-180m-flash! 8.537.78–9.34 31.9× 21.3× 0.20 GB CC-BY-4.0
11 Voxtral-Mini-4B-Realtime-2602 9.057.99–10.18 0.9× 0.6× 4.73 GB Apache-2.0
12 nemotron-3.5-asr-streaming-0.6b 10.7810.02–11.57 14.5× 7.5× 0.70 GB OpenMDW-1.1
13 whisper-small 13.3012.47–14.19 12.1× 3.4× 0.25 GB Apache-2.0