A new leaderboard scores open text-to-speech models on objective metrics

A team on Hugging Face launched the Open TTS Leaderboard, a ranking for open text-to-speech and voice-cloning models that leans on objective metrics instead of waiting weeks for human votes to pile up. It scores three things automatically: intelligibility, by running generated audio back through Qwen3 ASR and measuring word and character error rates; speed, as a real-time factor for batched inference and time-to-first-audio for streaming, measured on both an H200 GPU and CPU; and speaker similarity, as the cosine distance between WavLM embeddings of the generated and reference voices. That turns an evaluation that used to take a couple of weeks into a couple of hours.

The point is coverage of open models. The maintainers note in the write-up that on competing arena-style leaderboards only 16 of 92 models have open weights, so most practitioners cannot actually run the top entries. Early results put Kokoro-82M, Supertonic-3, and fishaudio's s2-pro near the top for English, with kyutai's pocket-tts leading on streaming latency. A "Listen" tab still collects human votes on real samples, so the objective scores sit alongside preference rankings rather than replacing them.

Why it matters

If you need a speech model you can self-host, this gives you error rates and latency numbers for open options side by side, so you can match a model to your language and latency budget instead of guessing from demos.