Model for memory-based music transcription and its Variational Bayes solution

The problem of memory based music transcription is considered and a probabilistic model for polyphonic music is proposed. Parameters of the model correspond to labels of the pre-recorded sounds and their amplitudes. Since exact estimation of the parameters is computationally prohibitive, we develop an approximate estimation algorithm using the Variational Bayes approximation. Results of the proposed algorithm are compared to alternative algorithms on piano recordings.