Ant Group releases Ming-3.0 open source speech models

Ant Group’s BaiLing team has released the Ming-3.0 speech model series as open source. The release includes five models categorized into three functional areas: speech synthesis, audio understanding and dialogue, and speech recognition.

The Ming-3.0-tts series features a ‘director mode,’ allowing creators to specify emotional shifts, word stress, and pacing through text descriptions or reference audio clips. These models are already utilized within Ant Group for applications such as the Afu app, doctor voice cloning, and the Xiaohu office assistant.

The audio understanding models, including Ming-3.0-audio-flash and a smaller version, are designed to transcribe and summarize long recordings while tracking speaker identity, emotion, and environmental sounds. Finally, the Ming-3.0-asr-tiny model focuses on speech recognition, utilizing training corpora that include Mandarin, English, regional dialects, and industry-specific hotwords to improve accuracy in real-world business scenarios.