Detection and restoration of sound of flute embedded in noise using real-time Kalman filter

Abstract

The restoration of flute notes embedded in noise is formulated as a state-estimation problem of a dynamic system. A single Kalman filter, with a given state-transition matrix, is implemented in real-time to recover the corresponding note as well as some of the neighbouring notes. In order to restore a continuous piece of music played by flute, a bank of Kalman filters (with different state-transition matrices) can be used. Event detectors are employed to detect the change of underlying system model. The overall restored sound is the output of Kalman filter, whose model is valid at a given time instant. >