Audio Separation
We are developing new approaches and algorithms to solve the problem of source separation. Although we focus on the problem of audio mixtures our work is directly applicable to multiple types of signals ranging from audiovisual, to biomedical/chemical, to vibrations and others. Our focus is to enable processing on non-clean data to not be influenced by interfering sources.
Background & Objective: As is often the case audio signals are captured with interference from other sources. Since most time-series algorithms are designed to work on a single source signal, this interference results into suboptimal performance. This problem becomes especially prevalent when dealing with systems that perform speech recognition, or when users need to evaluate noisy data (e.g. a noisy cell phone recording). Our objective is to investigate methods with which we can perform various tasks on multiple sound recordings as well as we can with single sound ones.
Technical Discussion: We have recently presented a sequence of papers which describe some new techniques for source separation based on latent model decompositions. We have fine-tuned our techniques to work on source separation problems and our results are very competitive with the state of the art.
Future Direction: There are currently multiple ways to move ahead with our work and we are actively investigating multiple extensions and applications. We expect to make sever more contributions to the field in the coming year.
Contacts:
Bhiksha Raj
Ajay Divakaran
| Technical Reports: | |
| Convolutive Speech Bases and their Application to Supervised Speech Separation | |
| Latent Dirichlet Decomposition for Single Channel Speaker Separation | |
| Latent Variable Decomposition of Spectrograms for Single Channel Speaker Separation | |
Technology Area: Sensor and Data Systems
Modification Date: November 1, 2007
