The present study investigates ways in which the "timbre space" metaphor may be used in creative ways for instrumental composition. Numerous tools for concatenative sound synthesis share today the ability to represent in an ndimensional space large quantities of sound, thus displaying on a map data which originally unfolded in time. If the potential of such systems for creating interactive instruments is an evidence, their affordance as musical scores needs further assessment, for control over time becomes unknown territory. When porting to VR such representation of sonic data, the score becomes a 3-dimensional map (or world) in which the user typically navigates freely. Experimentation through composition, instrument design and improvisation have shown a potential of simulation of plausibly automatised acoustical instrument, using machine learning techniques to model virtual instruments out of relatively small quantities of data (e.g. 20 minutes of audio to model a clarinet). The method offers promising avenues for the exploration of instrumental fragments clustered by timbre, register, dynamic, instrumental techniques. Whether or not such maps identify as musical scores, they contribute to addressing a problem formulated by Lev Manovitch: "how to merge database and narrative into a new form".