handy models

speech-to-text benchmarks for transcribe.cpp, the engine behind Handy

the whole grid

Every model × every benchmarked language. Darker is worse. A ring means the model card claims the language but there is no benchmark. Click any cell for the language leaderboard.

error %: 0491317212630+ claimed, not benchmarked
arbgyuehrcsdanlenetfilfifrdeelhihuiditjakolvltmkmsmtzhnbfaplptroruskslessvthtrukvi whisper-large-v3 whisper-large-v3-turbo whisper-medium whisper-small Qwen3-ASR-1.7B Qwen3-ASR-0.6B Fun-ASR-MLT-Nano-2512 nemotron-3.5-asr-streaming-0.6b canary-1b-v2 parakeet-tdt-0.6b-v3 parakeet-primeline cohere-transcribe-03-2026 Voxtral-Mini-4B-Realtime-2602 SenseVoiceSmall canary-180m-flash Breeze-ASR-25 parakeet-unified-en-0.6b gigaam-v3-e2e-rnnt cohere-transcribe-arabic-07-2026

The 63 card-only languages (Whisper’s long tail) are not columns here — find them under languages → “on model cards only”.