In the field of AI speech generation, the open-source camp is welcoming a major breakthrough. The independent AI evaluation agency Artificial Analysis recently published the latest speech large model (TTS) Arena leaderboard, and Ming-3.0-tts, a brand-new open-source text-to-speech model launched by the Ant Group team, delivered an outstanding performance, taking first place in multiple core metrics in one sweep.
As a lightweight and efficient model with compact parameters, Ming-3.0-tts offers a variety of preset voices as well as stateless voice cloning capabilities. Its weights have been fully open-sourced on the Hugging Face platform under the MIT license, with the release of a 0.5B-parameter dense model version and a variant version, Ming-3.0-tts-tiny, which adopts a 7.9B Mixture of Experts (MoE) architecture.
In specific authoritative benchmark tests and sub-leaderboards, Ming-3.0-tts demonstrated strong technical dominance:
First, on the Pronunciation Robustness leaderboard, the model achieved a high score of 93.2%, tying for first place among the 31 mainstream models worldwide in the evaluation, on the same level as Fish Audio's Drama3Preview and ahead of Eleven v4's 91.7%. Especially when handling high-difficulty tasks such as expanding abbreviations and precise sequence retention, the model performed outstandingly, accurately handling the pronunciation challenges of all kinds of complex texts.
Second, in the Controlled Voice Arena—which conducts side-by-side blind tests using a unified set of 8 cloned voices—Ming-3.0-tts successfully topped the open-source model leaderboard, achieving an Elo score of 1,069 across 1827 evaluations, ahead of Drama3Preview's 1,059 and Breeze TTS2's 1,019, ranking 13th on the overall leaderboard.
In addition, on the Provider Voice leaderboard, the model achieved an Elo score of 1,188 across 1957 evaluations, ranking third among open-source models and 19th on the overall leaderboard. Overall, Ming-3.0-tts successfully claimed first place among open-source models in both key dimensions of the three major core speech leaderboards—"Controlled Voice" and "Pronunciation Robustness"—delivering an impressive scorecard to developers worldwide.
Project address: https://huggingface.co/spaces/hugging-apps/ming-3-tts-tiny