Interactive User-Feedback for Sound Source Separation

Copyright is held by the author/owner(s). IUI’13, March 19–22, 2012, Santa Monica, California, USA. This work was performed while interning at Adobe Research. Abstract Machine learning techniques used for single-channel sound source separation currently offer no mechanism for user-feedback to improve upon poor results and typically require isolated training data to perform separation. To overcome these issues, we present work that applies interactive machine learning principles to incorporate continual user-feedback into the source separation process. In particular, we allow end-users to annotate errors found in source separation estimates by painting on time-frequency displays of sound. We then employ a posterior regularization technique to make use of the annotations to obtain refined source separation estimates and repeat the process until satisfied. An initial prototype shows that the proposed method can significantly improve separation quality compared to previous work and facilitate separation without isolated training data.