Ant Group has open-sourced its Ming-3.0 family of speech models, released on October 20. The release makes multiple models available for different tasks, including speech synthesis, audio understanding and conversation, and speech recognition. Developers can use individual models or combine them when building voice-based applications.
The text-to-speech components include voice cloning from reference recordings and “voice design” driven by natural-language descriptions. One source says the TTS models are published under the MIT license in two sizes, and provide fine control features such as a set of paralinguistic tags that influence speech characteristics. The release also supports multiple languages, including major options plus regional Chinese varieties, and can be used with multi-speaker podcast-style content.
Across the two outlets, the emphasis is on access for developers and controllable, voice-capable capabilities rather than any new commercial product launch. Both describe the models as a toolkit for constructing speech and conversation systems.