Multipitch Estimation Using Sparse Impulse Distributions and Instrument Specific Priors
Gautham J. Mysore and Paris SmaragdisWe present an algorithm for the concurrent estimation of the relative pitch tracks of multiple instruments in a sound mixture. A relative pitch track and a timbral signature is concurrently estimated for each instrument in the mixture. The estimation is carried out by performing multiple simultaneous deconvolutions using shift-invariant probabilistic latent component analysis on a time-frequency magnitude representation (constant-Q transform) of the sound mixture. The deconvolutions yield a kernel distribution (spectral signature) and impulse distribution (from which the relative pitch track is extracted) for each instrument in the mixture. Entropic prior distributions are used on the impulse distributions in order to make them sparse. Spectral characteristics of certain instruments are specified using instrument specific prior distributions on the kernel distributions. Additionally, a Kalman filter type smoothing is used to enforce temporal continuity of pitch tracks.