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

Arabic · 17 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 cohere-transcribe-arabic-07-2026!overall pickaccurate pick 11.06 8.3× 4.3× 2.25 GB Apache-2.0
2 cohere-transcribe-03-2026! 13.60 8.1× 4.3× 2.25 GB Apache-2.0
3 Voxtral-Mini-4B-Realtime-2602 14.53 0.8× 0.6× 4.41 GB Apache-2.0
4 Qwen3-ASR-1.7B 14.91 4.1× 2.5× 2.04 GB Apache-2.0
5 whisper-large-v3 14.92 2.0× 1.0× 1.55 GB Apache-2.0
6 whisper-large-v3-turbo 15.48 3.1× 1.4× 0.83 GB Apache-2.0
7 nemotron-3.5-asr-streaming-0.6bfast pick 15.93 17.1× 10.6× 0.70 GB OpenMDW-1.1
8 whisper-large-v2 17.06 2.0× 1.0× 1.55 GB Apache-2.0
9 whisper-large 19.47 2.0× 1.0× 1.55 GB Apache-2.0
10 whisper-medium 21.90 4.1× 1.9× 0.77 GB Apache-2.0
11 Qwen3-ASR-0.6B 24.51 9.7× 5.3× 0.79 GB Apache-2.0
12 moonshine-base-ar 24.62 17.9× 11.0× 0.07 GB MIT
13 Fun-ASR-MLT-Nano-2512! 25.79 8.8× 5.5× 0.83 GB FunASR Model Open Source License Agreement v1.1
14 moonshine-tiny-ar 26.70 29.6× 19.7× 0.03 GB MIT
15 whisper-small 32.15 11.7× 5.9× 0.25 GB Apache-2.0
16 whisper-base 52.74 29.0× 19.3× 0.08 GB Apache-2.0
17 whisper-tiny 67.48 43.1× 39.8× 0.04 GB Apache-2.0