Ant Group's Ming-3.0-tts led open-weights models in Controlled Voice and tied for first overall in Pronunciation Robustness, Artificial Analysis reported Oct. 20. Its Provider Voice standing was lower: third among open-weights models and 19th overall, according to the evaluator.
The voice tests use different setups. In Controlled Voice, every model is compared using the same eight cloned voices. Artificial Analysis gave Ming a 1,069 Elo rating across 1,827 appearances, placing it first among open-weights models but 13th overall. In Provider Voice, each model uses its own native voices. There, Artificial Analysis reported a 1,188 Elo rating across 1,957 appearances for Ming.
For Pronunciation Robustness, human reviewers judged clips of challenging text in four categories against pre-agreed accepted pronunciations. Artificial Analysis reported a 93.2% score for Ming, tying Fish Audio's Drama 3 Preview for first among 31 models; Eleven v4 scored 91.7%. Ming led two of the four categories, Expanding Shorthand at 96.5% and Preserving Exact Sequences at 90.6%. It scored 91.0% in Contextually Appropriate and 88.5% in Standalone Terms, according to the evaluator.
Artificial Analysis describes Ming-3.0-tts-0.5B as a dense 0.5B-parameter model with preset voices and stateless voice cloning. It says the weights are available on Hugging Face under the MIT License, alongside a 7.9B mixture-of-experts variant called Ming-3.0-tts-tiny.
