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

Portuguese · 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 3.883.45–4.37 2.1× 0.6× 1.55 GB Apache-2.0
2 whisper-large-v3-turbo 4.173.73–4.69 3.4× 0.8× 0.83 GB Apache-2.0
3 Qwen3-ASR-1.7B 4.373.95–4.83 3.8× 2.0× 2.04 GB Apache-2.0
4 canary-1b-v2!overall pickfast pick 4.504.04–5.01 13.2× 6.7× 1.10 GB CC-BY-4.0
5 parakeet-tdt-0.6b-v3 4.964.49–5.52 12.5× 7.5× 0.72 GB CC-BY-4.0
6 whisper-medium 5.074.61–5.57 4.3× 1.1× 0.77 GB Apache-2.0
7 parakeet-primeline 5.174.72–5.66 12.5× 7.5× 0.72 GB CC-BY-4.0
8 cohere-transcribe-03-2026! 5.184.71–5.66 8.0× 3.0× 2.41 GB Apache-2.0
9 Voxtral-Mini-4B-Realtime-2602 5.234.60–5.94 0.9× 0.6× 4.73 GB Apache-2.0
10 Qwen3-ASR-0.6B 6.576.05–7.12 8.0× 4.3× 0.79 GB Apache-2.0
11 whisper-small 7.656.96–8.31 12.1× 3.4× 0.25 GB Apache-2.0
12 nemotron-3.5-asr-streaming-0.6b 8.527.77–9.34 14.5× 7.5× 0.70 GB OpenMDW-1.1
13 Fun-ASR-MLT-Nano-2512! 28.2426.85–29.59 9.0× 4.5× 0.83 GB FunASR Model Open Source License Agreement v1.1