What is Timbre?

Michèle Castellengo, LAM

We know that timbre is a perceptual notion that only listeners are able to catch, describe and evaluate. Recently it has become possible to use computers to automatically recognize the sound of musical instruments with good performance, which seems paradoxical. The question then arises: do musicians and scientists refer to the same notion when they use the word timbre? Is it necessary to take into account different forms of timbre conceptualization? After a short survey of timbre studies carried out both by scientists and by musicians, we will show that timbre is, at least, a two-pronged notion, the first referring to sound identification, the second to sound qualification. With the help of sound examples we will propose a general framework for sound perception in which timbre listening is only a specialized behavior applied to musical sounds, a view confirmed by recent  neuroscience research.

Download slides »

How do we perceive timbre?

Timothy D. Griffiths, Newcastle University

I will describe a framework for the representation of acoustic time and frequency structure in the primate brain based on functional imaging in the ascending pathway and auditory cortex. These mechanisms constitute initial sensory representation after which timbre is abstracted based on a combination of cues related to the time and frequency structure of sound. We carry out an approach to timbre in which synthetic stimuli allow the identification of networks for the representation of timbral perceptual dimensions in non-core auditory cortex. Dynamic Casual Modelling extends network identification by supporting constructive systems for timbre similar to those that have also been suggested for pitch. Finally I will describe an acquired disorder of timbre perception that might be called ‘dystimbria’ and consider how this can be understood with respect to the normal systems.

Download slides »

Deep Network Geometry of Timbre

Stéphane Mallat, ENS

Deep convolution networks are the new winners of many audio classification problems. These architectures provide geometric signal representations, which are connected to classical audio features such as MFCC and modulation spectrum, but which can also go well beyond. There are similarities with image geometry, but audio signals have different geometrical properties. Geometric timbre representations are studied through analysis-synthesis audio experiments, which raise many open issues, also related to physiological cortical models.

Download slides »