The authors introduce the idea of performing it Intelligent ICA to focus on and separate a specific instrument, voice or sound source of interest. This is achieved by incorporating high-level probabilistic priors in the ICA model that characterise each instrument or voice. For instrument modelling, we experimented with various feature sets previously used for instrument or speaker recognition. Prior training of a Gaussian Mixture Model for each instrument was performed. The order of the feature vector, the number of gaussian mixtures and the training audio data length were kept to reasonably minimum levels.
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member and would like to subscribe to the E-Library then Join the AES!
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.