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

Dutch · 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 whisper-large-v3most accurate 5.424.81–6.07 2.1× 0.6× 1.55 GB Apache-2.0
2 whisper-large-v3-turbo 5.985.35–6.64 3.4× 0.8× 0.83 GB Apache-2.0
3 canary-1b-v2!overall pickfast pick 6.285.63–6.98 13.2× 6.7× 1.10 GB CC-BY-4.0
4 cohere-transcribe-03-2026! 7.166.15–8.15 8.0× 3.0× 2.41 GB Apache-2.0
5 Qwen3-ASR-1.7B 7.436.73–8.15 3.8× 2.0× 2.04 GB Apache-2.0
6 parakeet-tdt-0.6b-v3 7.666.83–8.46 12.5× 7.5× 0.72 GB CC-BY-4.0
7 parakeet-primeline 8.497.76–9.26 12.5× 7.5× 0.72 GB CC-BY-4.0
8 Voxtral-Mini-4B-Realtime-2602 8.887.90–10.02 0.9× 0.6× 4.73 GB Apache-2.0
9 whisper-medium 10.409.48–11.33 4.3× 1.1× 0.77 GB Apache-2.0
10 nemotron-3.5-asr-streaming-0.6b 13.6112.47–14.70 14.5× 7.5× 0.70 GB OpenMDW-1.1
11 Qwen3-ASR-0.6B 13.9012.88–14.93 8.0× 4.3× 0.79 GB Apache-2.0
12 whisper-small 18.4817.29–19.57 12.1× 3.4× 0.25 GB Apache-2.0
13 Fun-ASR-MLT-Nano-2512! 42.9741.38–44.61 9.0× 4.5× 0.83 GB FunASR Model Open Source License Agreement v1.1