Open TTS Leaderboard adds scalable multilingual model evaluation
A new Open TTS Leaderboard evaluates text-to-speech systems with objective measures intended to scale faster than human-voting arenas. The project says the Hugging Face Hub had more than 8K TTS models as of September 30, 2026, while evaluation remained fragmented. Its metrics cover intelligibility through word and character error rates, inference speed through RTFx and time-to-first-audio, and speaker similarity using WavLM embeddings.
The leaderboard supports multilingual comparisons and voice-cloning tests. Its default rankings use macro-average WER on English portions of Seed TTS Eval and CV3 Eval, while other language views use available CV3 Eval data; CER is reported for Chinese, Japanese and Korean. hexgrad/Kokoro-82M, Supertone/supertonic-3 and fishaudio/s2-pro lead the combined English WER view. k2-fsa/OmniVoice, fishaudio/s2-pro and FunAudioLLM/Fun-CosyVoice3-0.5B-2512 are identified as strong multilingual models.
A Listen tab lets users compare outputs and submit feedback, while a Streaming tab ranks time-to-first-audio on identical prompts and hardware. The creators stress that automated scores do not measure naturalness or listener preference and therefore do not replace human rankings. Evaluation scripts are planned for open-sourcing.
Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.