StepAudio 3 Music

StepAudio 3 Music is a high-fidelity music generation model developed by StepFun. Part of the StepAudio 3 family of models, it is designed to synthesize complete musical compositions including vocals, harmony, and complex instrumental arrangements from natural language descriptions. The model follows the architectural lineage of the ACE-Step research project, utilizing a hybrid system that combines a large-scale language model as a "musical planner" with a Diffusion Transformer (DiT) as the audio engine.

The model excels in generating music with coherent structural transitions, supporting segments such as verses, choruses, and bridges. It is capable of handling lyrics in over 50 languages and provides fine-grained control over musical parameters including genre, emotional tone, and tempo. On industry leaderboards, StepAudio 3 Music has demonstrated high performance in both instrumental and vocal categories, competing with specialized commercial and open-source alternatives.

Key features include high-speed synthesis—often generating a full-length track in under 20 seconds on enterprise hardware—and the ability to maintain long-range melodic and rhythmic consistency. The model supports advanced editing capabilities such as repainting (audio inpainting) and vocal-to-BGM conversion, allowing producers to manipulate specific elements of a track while preserving the original style. Users can guide the generation process by providing explicit lyrics and detailed genre tags, or allow the model to brainstorm song blueprints through its internal reasoning capabilities.