Closing the Loop of Sound Evaluation and Design (CLOSED)

How should an incoming email sound? Or a mouse click, a closing door, a Harley Davidson motorcycle? Answering this question is the goal of the emerging discipline of sound design. Of special importance is the development of machine learning based tools that can predict the perceived quality of everyday sound automatically. Is this sound an opening or a closing door? Can the algorithm "hear" if the door is made of wood, metal, or glass? Does the slamming of the door sound agressive?

Everyday sounds can be described on several levels: basic properties of the object (material, thickness, extent, force), a temporal patterning (accelerating, decelerating, fast, slow, irregular rhythm), and its function (e.g. opening/closing door). Sounds have to be labelled with these attributes first. Then the sound is processed frame-by-frame. For each frame, feature vectors are calculated describing the sound's timbre, pitch, and onsets. Furthermore the sound characteristics are defined by the sequel of basic events, e.g. for a door: pressing the handle, opening the door, squeaking. Therefore, we need an algorithm to explore the temporal structure of the sequel of feature vectors. The final aim of this project is the development of tools for sound design. The work is done in cooperation with psychologists, sound designers and experts on sound synthesis. Read more: http://closed.ircam.fr