Ming-3.0-tts

Ming-3.0-tts is an open-source text-to-speech model family developed by Ant Group's inclusionAI team, featuring dense and mixture-of-experts variants designed for fine-grained voice control and high-fidelity speech generation. The model family spans a 0.5B dense architecture built on Qwen2.5-0.5B and a 7.9B-A1.3B MoE variant utilizing a Ling V3 backbone, supporting multilingual synthesis across seven languages and regional dialects like Sichuanese and Cantonese.

Architecture and Capabilities

The architecture integrates a language model backbone with a diffusion transformer (DiT) patch head and an AudioVAE decoder that generates 44.1 kHz audio waveforms. Ming-3.0-tts introduces advanced capabilities such as zero-shot voice cloning from short reference clips, natural-language voice design for creating custom speaker personas from text prompts, and precise delivery controls including explicit emotional, pitch, and speed adjustments. It also incorporates inline paralinguistic tags for natural acoustic events like sighs, laughter, and breath sounds.

Usage Guidelines

The model accepts instructions in natural language for style direction and supports multi-speaker dialogue setups reminiscent of podcast workflows. Text input is normalized using specialized grammars, ensuring consistent pronunciation across numbers, units, and symbols. Designed for robust audio generation, the models offer fine control over vocal delivery while preserving flexibility across diverse speech applications.