Separating desired audio content from undesired content

An audio processing method comprising:

• training a first dictionary for clean desired content;
• receiving an audio signal to be processed, and providing a multi-dimensional non-negative representation of the received audio signal;
• learning a second dictionary for undesired content;
• decomposing the multi-dimensional non-negative signal representation into desired content and undesired content;
• wherein the desired content is represented by the first dictionary and the undesired content is represented by the second dictionary, and wherein the multi-dimensional non-negative signal representation has at least two dimensions and includes a forward and an inverse transform;
• imposing sparsity on activations of basic elements of the first or the second dictionary, wherein a degree of sparsity is controlled by setting a minimum number of components with significant activations of said first or second dictionary;
• and
• separating the desired content from the undesired content.

A system comprising:

• one or more processors;
• and
• a non-transitory computer-readable medium storing instructions that, when executed by the one or more processors, cause the one or more processors to perform an audio processing method comprising:
• training a first dictionary for clean desired content;
• learning a second dictionary for undesired content;
• decomposing a multi-dimensional non-negative signal representation of a received audio signal into desired content and undesired content by performing convolutional non-negative matrix factorization (CNMF) in multiple layers, each layer having a respective non-negative matrix representation;
• wherein the desired content is represented by the first dictionary and the undesired content is represented by the second dictionary, and wherein the multi-dimensional non-negative signal representation has at least two dimensions and includes a forward and an inverse transform;
• and
• separating the desired content from the undesired content.

A non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform an audio processing method comprising:

• training a first dictionary for clean desired content;
• learning a second dictionary for undesired content;
• decomposing a multi-dimensional non-negative signal representation of a received audio signal into desired content and undesired content by performing convolutional non-negative matrix factorization (CNMF) in multiple layers, each layer having a respective non-negative matrix representation;
• wherein the desired content is represented by the first dictionary and the undesired content is represented by the second dictionary, and wherein the multi-dimensional non-negative signal representation has at least two dimensions and includes a forward and an inverse transform;
• and
• separating the desired content from the undesired content.