Prior Subspace Analysis for Drum Transcription

This paper introduces the technique of Prior Subspace Analysis (PSA) as an alternative to Independent Subspace Analysis (ISA) in cases where prior knowledge about the sources to be separated is available. The use of prior knowledge overcomes some of the problems associated with ISA, in particular the problem of estimating the amount of information required for separation. This results in improved robustness for drum transcription purposes. Prior knowledge is incorporated by use of a set of prior frequency subspaces that characterise features of the sources to be extracted. The effectiveness and robustness of PSA is demonstrated by its use in a simple drum transcription algorithm. 1. INDEPENDENT SUBSPACE ANALYSIS Independent Subspace Analysis (ISA) provides a means of attempting sound source separation from single channel mixtures [1]. Based on redundancy reduction techniques, it represents sound sources as low dimensional independent subspaces in the timefrequency plane. To carry out ISA the single channel mixture signal is converted to a time-frequency representation such as a spectrogram. It is then assumed that the overall spectrogram Y results from the superposition of a number of unknown statistically independent spectrograms Yj, yielding: ∑ = = l