The Open TTS Leaderboard trades votes for instruments
Hugging Face's Open TTS Leaderboard ranks open-source, multilingual text-to-speech and voice cloning models on objective metrics: ASR-based error rate, speaker similarity and latency. Results take hours instead of weeks of voting. It's an evaluation resource, not a new capability, and it says plainly what it can't measure.
Hugging Face has launched the Open TTS Leaderboard, a benchmark for open-source text-to-speech and voice cloning models that covers many languages and uses objective metrics. It isn't a new model or a new capability. It is a shared, repeatable place to compare models, which the open speech community has lacked while releases piled up: the announcement counts more than 8K TTS models on the Hub as of September 30, 2026.
What shipped
Every model is scored on three things:
- Intelligibility. Word or character error rate (WER/CER) between the input text and a transcript of the generated audio from Qwen3 ASR, which the authors call the top open-source model on the Open ASR Leaderboard.
- Speaker similarity. Cosine similarity between WavLM speaker embeddings of the output and the reference clip, when voice cloning is on.
- Speed. Inverse real-time factor (RTFx) for batched inference on an H200 GPU, and time-to-first-audio (TTFA) at batch size 1 on GPU and CPU.
The English test sets are Seed TTS Eval (paper) and CV3 Eval zero-shot (paper). Other languages can be switched on, with Chinese, Japanese and Korean scored by CER. Pareto plots show error rate against throughput and model size, so you see the tradeoff rather than a single ranked list.
A Listen tab plays the audio behind the scores and takes votes from signed-in users. A Streaming tab ranks models by median TTFA over 50 fixed English prompts from CV3 Eval. For a model that can't stream, TTFA is the time to generate the whole clip, because playback can't start any earlier. That's the right definition.
Why it matters
Arenas such as TTS Arena v2 rank models by pairwise human votes. Human preference is still the best judge, but arenas move at the speed of voting, and they lean commercial: per the announcement, only 16 of the 92 models on Artificial Analysis's arena were open-weights as of September 30, 2026. Adding an API model takes a key; adding an open model means someone has to host and serve it. The authors say objective metrics cut evaluation from a couple of weeks of vote collection to a couple of hours. For open models, that's the difference between showing up on a leaderboard and not.
The multilingual default matters as much. English results don't carry over. hexgrad/Kokoro-82M, Supertone/supertonic-3 and fishaudio/s2-pro lead on English WER, while k2-fsa/OmniVoice, fishaudio/s2-pro and FunAudioLLM/Fun-CosyVoice3-0.5B-2512 stand out on the multilingual view. If you're shipping voice outside English, check the right view before you build.
My take
The release is honest about its limits. ASR-based WER stands in for intelligibility and speaker similarity for voice identity; neither captures naturalness, expressiveness or preference, which is why the Listen tab is part of the design. I'd add one caution: the ASR judge is part of the measurement. In languages where it is weaker, small gaps between models say something about the judge as well as the TTS. And any "best open TTS" claim should name the language, the dataset and whether cloning was on.
For teams building voice agents, the Streaming tab is the practical part, since TTFA at batch size 1 is the number users feel. The evaluation scripts aren't public yet; the authors say they'll open-source them, as they did for the Open ASR Leaderboard. Until then, I treat the board as a strong filter for deciding which models deserve human listening tests.