Abstract
We present a new method for score-informed source separation, combining ideas from two previous approaches: one based on paramet- ric modeling of the score which constrains the NMF updating process, the other based on PLCA that uses synthesized scores as prior probability distributions. We experimentally show improved separation results using the BSS EVAL and PEASS toolkits, and discuss strengths and weaknesses compared with the previous PLCA-based approach.