Improvements of continuous model for memory-based automatic music transcription

Automatic music transcription is a process recovering the most likely combination of sounds that produced the recorded audio signal. We are concerned with memory-based approach, where the observed signal is modeled as a superposition of sounds from a library. Moreover, we assume that only parts of the sounds can be played. The number of possible combinations is excessive and exact estimation is computationally prohibitive. We propose to transform the original discrete-event model into a less restricted parametrization and impose the constraints in a soft way via prior information. The resulting model is a non-linear state-space model with Gaussian disturbances. The posterior estimates are evaluated by the extended Kalman filter. Performance of the model is studied in simulation and it is shown that it outperforms previously published methods.