Audio source separation by basis function adaptation

The problem of audio source separation from a monophonic sound mixture having known instrument types but unknown timbres is presented. An improvement to the Probabilistic Latent Component Analysis (PLCA) source separation method is proposed. The technique uses a basis function dictionary to produce a first round PLCA source separation. The PLCA weights are then refined by incorporating note onset information. The source separation is then performed using a second round PLCA in which the refined weights are held fixed, and the basis functions are updated. Preliminary experimental results on mixtures of two instruments are quite promising, showing a 6 dB improvement in SIR over standard PLCA.

• Type: Journal
• Year: 2011
• Where: ICASSP
• Authors: Yinyi Guo, Mofei Zhu
• Keywords: Basis Functions, ICASSP 2011, Signal Processing, Source Separation, Source Separation Method