Ant Group releases open-source Ming-3.0 speech models

Ant Group announced the release of its Ming-3.0 model family on October 20, providing five distinct models for speech synthesis, audio understanding, and speech recognition. The release is designed for developers, who can deploy the models individually or integrate them to create custom voice-based applications.

The text-to-speech component supports voice cloning from reference samples and voice generation via natural language prompts. Users can further customize output by adjusting pacing, emphasis, and emotional tone. Additionally, the models perform transcription and summarization tasks on long-form audio while preserving data regarding speaker identity and ambient sounds.

Ant Group has already implemented the technology within its internal ecosystem, including the Afu health application and Alipay search functions. By moving these tools to an open-source model, the company joins a broader industry trend of making specialized speech AI accessible to independent developers.