This project creates a synthesis engine that predicts the timbre of arbitrary acoustic instruments. Musically meaningful, it is controlled by the perceptual features extracted from the audio stream of an acoustic or electric violin in real-time. The timbre models are built from the analysis of pitch, loudness, brightness, and the instantaneous power spectrum of real recordings. Although highly compressed, the sound quality is fully preserved. Additionally, the playability remains smooth, expressive, predictable, and adequate for sound morphing.