VIOLET: Violin Expression Rendering with Controllable Articulation and Dynamics

Expressive music synthesis for continuously excited instruments poses a fundamentally different problem from percussive instruments like piano, as realistic rendering depends on both note-level playing techniques and continuously varying dynamics. In music production, high-quality violin renderings are typically achieved with large sample libraries and labor-intensive MIDI programming of technique specification and continuous controllers. Bridging symbolic programming and realistic violin rendering therefore requires datasets and generative models that represent these controls explicitly. We present VIOLET, a latent-diffusion framework for controllable violin synthesis, together with CSV-TD, a new 48 kHz synthetic dataset of violin audio aligned with MIDI notes, note-level technique labels, and continuous dynamics control curves. VIOLET operates in a compressed audio latent space using a DiT backbone trained with rectified flow, and synthesizes high-fidelity audio conditioned on notes, techniques, and continuous control signals. Objective and subjective evaluations show that the proposed method produces natural and expressive violin performances with near-perfect technique adherence, accurate pitch and timing alignment, and reflects dynamics control. The system outperforms prior violin synthesis approaches and achieves audio quality comparable to commercial virtual instruments, narrowing the gap between symbolic programming and realistic acoustic rendering.