Voice Bench
Speech in 70+ languages · one 8-core CPU server · no GPU
CPU –RAM –– hot

Every language below runs on this one server: speech → text (ASR) and text → speech (TTS). Error rates are measured on held-out real recordings and on arbitrary text (numbers, units, names), and graded against the best open models and the leading commercial services in each language.

CER<22–55–10≥10 WER<55–1010–20≥20 gradeaheadon parclosebehindfar behind

CER / WER = share of characters / words wrong, lower is better. ASR: our model on real speech. TTS: what listeners' speech recognisers understand of our voice reading arbitrary text. Grades compare with the best open model and the best commercial service we tested in that language (“no commercial” = none supports it; “not run” = not benchmarked). Cost columns: measured on CPU, 8 threads — RAM of the loaded model, latency for a 3-second utterance (ASR) / to the first audio (TTS), and how many times faster than real time it runs.

Language

Text → speech

Speech → text

Real recordings — press one to transcribe it, or upload your own.

Speak in
Hold to speak · release to hear it back
Model memory

A hot model sits in RAM and answers at once. A cold one is loaded on demand — in seconds — and the least recently used model is released when memory runs short.

    CPU
    Memory
    Disk read / write