centre for digital music Blind Audio Source Separation

Most audio signals are mixtures of several audio sources which are active simultaneously. For example, live debates are mixtures of several speakers, music CDs are mixtures of musical instruments and singers, and movie soundtracks are mixtures of speech, music and natural sounds. Blind Audio Source Separation (BASS) is the problem of recovering each source signal from a given mixture signal. This report provides a tutorial review of established and recent BASS methods as applied to the separation of realistic audio mixtures, focusing on situations where large microphone arrays or other unusual microphone arrangements are not available. Our first goal is to show that a large range of assumptions can be made to separate a given mixture signal. Our second goal is to emphasize the importance of audio-specific issues in the design of BASS algorithms. Thus we consider the BASS problem in its full generality and we point out the modeling assumptions and the limitations of each class of algorithms. For the sake of clarity, we describe approaches relating to different historical viewpoints within a general statistical framework. We do not discuss implementation details nor other problems related to BASS such as dereverberation and remixing.