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 CER — lower is better. Speed is × realtime on Handy target hardware.

accuracy vs speed

Japanese · 20 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 Fun-ASR-MLT-Nano-2512!overall pickaccurate pick 2.32 8.8× 5.5× 0.83 GB FunASR Model Open Source License Agreement v1.1
2 whisper-large-v3 4.81 2.0× 1.0× 1.55 GB Apache-2.0
3 whisper-large-v3-turbo 4.82 3.1× 1.4× 0.83 GB Apache-2.0
4 cohere-transcribe-03-2026! 5.13 8.1× 4.3× 2.25 GB Apache-2.0
5 Qwen3-ASR-1.7B 5.29 4.1× 2.5× 2.04 GB Apache-2.0
6 Voxtral-Mini-4B-Realtime-2602 5.44 0.8× 0.6× 4.41 GB Apache-2.0
7 whisper-large-v2 5.56 2.0× 1.0× 1.55 GB Apache-2.0
8 granite-speech-4.1-2b! 6.30 4.6× 2.8× 2.38 GB Apache-2.0
9 granite-4.0-1b-speech! 6.44 4.6× 2.8× 2.38 GB Apache-2.0
10 whisper-large 6.95 2.0× 1.0× 1.55 GB Apache-2.0
11 whisper-medium 7.35 4.1× 1.9× 0.77 GB Apache-2.0
12 SenseVoiceSmallfast pick 7.63 27.7× 19.7× 0.24 GB model-license (FunASR MODEL_LICENSE)
13 Fun-ASR-Nano-2512! 8.50 10.7× 7.9× 0.83 GB FunASR Model Open Source License Agreement v1.1
14 Qwen3-ASR-0.6B 8.61 9.7× 5.3× 0.79 GB Apache-2.0
15 moonshine-base-ja 11.11 23.4× 14.9× 0.07 GB MIT
16 whisper-small 12.81 11.7× 5.9× 0.25 GB Apache-2.0
17 moonshine-tiny-ja 13.44 41.2× 27.0× 0.03 GB MIT
18 nemotron-3.5-asr-streaming-0.6b 13.52 17.1× 10.6× 0.70 GB OpenMDW-1.1
19 whisper-base 25.28 29.0× 19.3× 0.08 GB Apache-2.0
20 whisper-tiny 39.38 43.1× 39.8× 0.04 GB Apache-2.0