Hugging Face launches an open TTS leaderboard

Hugging Face launched an open TTS leaderboard covering open-source and multilingual text-to-speech and voice cloning. The motivation is scale: more than 8,000 TTS models were on the Hub as of September 30, 2026, and only 16 of the 92 models on Artificial Analysis are open-weight, so picking one has been guesswork.

It scores three objective metric families. Intelligibility is measured as word or character error rate using Qwen3-ASR-1.7B. Speed covers inverse real time factor batched offline on H200, plus time to first audio for streaming at batch size one on both H200 and CPU. Speaker similarity is the cosine similarity of WavLM embeddings. Hugging Face says evaluation time drops from a couple of weeks of vote collection to a couple of hours.

On English word error rate the leaders are Kokoro-82M, Supertonic 3 and Fish Speech S2 Pro. On multilingual performance the leaders are OmniVoice, S2 Pro and Fun-CosyVoice3. Streaming leaders include Pocket-TTS, which performs well on both GPU and CPU.

The result most likely to change behaviour is that cloning can improve word error rate for some models, with Higgs TTS 3 4B and VoxCPM2 named, which is the opposite of the assumption that reference audio only adds noise.

The limitations are important and Hugging Face states them first. This does not replace human preference ranking. Word error rate is only an intelligibility proxy, and speaker similarity only estimates whether voice identity survived. Neither measures naturalness, expressiveness or listener preference, which is the thing people actually care about. The evaluation scripts are not open sourced yet, CPU results cover only a small and growing set of models, and Seed-TTS Eval has audio only for English and Chinese, so other languages score zero-shot on CV3 Eval only. There is a Listen tab for A to B comparison, with community voting behind a login to limit spam.

(Original source: Hugging Face)