@Ryrenz: So smooth—Ming-3.0-tts can clone a voice directly from a reference audio clip, and you can even tell it how to speak with a single sentence. It was just open-sourced yesterday by Ant Group’s inclusionAI. On Artificial Analysis’s October speech synthesis leaderboard, Ming-3.0-tts’s…

🔊 So smooth: Ming-3.0-tts can clone a voice directly from a reference audio clip, and you can tell it how to speak with a single sentence inclusionAI, part of Ant Group, just open-sourced it yesterday. On Artificial Analysis’s October speech synthesis leaderboard, Ming-3.0-tts ranked first overall for pronunciation robustness with 93.2% accuracy, and first among open-source models for voice controllability. When making voiceovers and audio content, the most frustrating thing isn’t a voice that doesn’t sound quite right—it’s mispronunciation. Whether abbreviations should be expanded, whether model numbers and identifiers can be read out exactly—in the past, you had to test each sentence, listen for problems, then go back and edit the text. Ming-3.0-tts scored 96.5% and 90.6% on these two measures, respectively, topping the leaderboard in both and saving those repeated rounds of trial and error. It also has 14 built-in paralinguistic tags, so you can add laughter, sighs, and similar effects simply by writing them into the text. Besides Mandarin, English, German, Spanish, French, Korean, and Japanese, it also speaks Sichuanese and Cantonese. The release includes a 0.5B dense model and a tiny mixture-of-experts model with 7.9B total parameters and 1.3B active parameters. The latter can also generate multi-speaker conversations from a script. The technical report isn’t out yet, but the inference code and evaluation tables are already in the repository. Getting the pronunciation right is the truly hard part of speech synthesis. GitHub: