Hugging Face launches Open TTS Leaderboard ranking open-source multilingual TTS and voice-cloning models on objective metrics

Hugging Face launched the Open TTS Leaderboard, a public ranking of open-source text-to-speech and voice-cloning models. It uses automated metrics only: word and character error rates from a Qwen3 speech recogniser on the Seed TTS Eval and CV3 Eval datasets, speed (RTFx on an H200 GPU, time to first audio on H200 and CPU) and speaker similarity via WavLM. Early English leaders on error rate are Kokoro-82M, supertonic-3 and s2-pro. OmniVoice, s2-pro and Fun-CosyVoice3-0.5B lead the multilingual table. The evaluation scripts are promised as open source "soon", and no links to them were supplied.

Developers choosing an open TTS model get a cheap, repeatable way to shortlist candidates on accuracy, latency and voice-cloning fidelity, including CPU latency that matters for self-hosting. It also gives open models a showcase next to arenas dominated by commercial APIs. The impact is narrow: it helps with screening, not with picking a voice people will enjoy listening to. Its value also depends on whether the scripts are actually released so others can reproduce the numbers.

The framing that human-preference arenas "can't scale" and that objective metrics cut evaluation from weeks to hours is partly self-serving, since a low error rate through an ASR model says little about naturalness, prosody or emotion, as the authors themselves admit. Rankings from one ASR judge (Qwen3) may favour models that sound clean to that recogniser, and the top positions reflect two test sets, not real-world use. The vote and Listen features need a Hugging Face login and have no stated volume yet.

1. 01Whether the evaluation scripts are open-sourced, and on what date, so third parties can reproduce the WER, RTFx and speaker-similarity figures.
2. 02How many models are listed after the first month and whether major releases (including commercial or API-only ones) are added, versus a static list of a few dozen.
3. 03Whether the objective rankings agree with human votes in the Listen tab and with TTS Arena v2, and whether model authors begin citing or tuning to the leaderboard.