Timbre Recognition with Combined Stationary and Temporal Features
Shlomo Dubnov, Xavier Rodet · University of Michigan Library Repository · 1998
In this paper we consider the problem of modeling spectro-temporal behaviour of musical sounds, with applications for musical instrument recognition. Using instanteneous sound features, such as cepstral envelopes and cepstral derivatives, the temporal evolution of the sound is transcribed into a new representation as a sequence of spectral features. Applying information-theoretic sequence matching methods, the sound dynamics can be modeled and compared. Testing this method on recordings of several solo musical pieces of di erent instrument, excellent matching results were obtained.