Real-Time Interactive Music Generation via Data-Free Streaming Consistency Distillation

The paper presents a novel framework for real-time interactive music generation that leverages a streaming autoregressive latent space, allowing for low-latency performance without the need for paired audio-latent datasets. It introduces music-aware consistency objectives to maintain acoustic fidelity, achieving a low real-time factor through parameter-efficient adaptation. This approach transforms generative music models into responsive instruments capable of integrating dynamic human inputs seamlessly, enhancing the potential for live human-AI collaboration in music creation.