Blind Speech Separation / Signals and Communication Technology (PDF)

This is the first book to provide a cutting edge reference to the fascinating topic of blind source separation (BSS) for convolved speech mixtures. Through contributions by the foremost experts on the subject, the book provides an up-to-date account of research findings, explains the underlying theory, and discusses potential applications. The individual chapters are designed to be tutorial in nature with specific emphasis on an in-depth treatment of state of the art techniques.

Blind Speech Separation is divided into three parts:

Part 1 presents overdetermined or critically determined BSS. Here the main technology is independent component analysis (ICA). ICA is a statistical method for extracting mutually independent sources from their mixtures. This approach utilizes spatial diversity to discriminate between desired and undesired components, i.e., it reduces the undesired components by forming a spatial null towards them. It is, in fact, a blind adaptive beamformer realized by unsupervised adaptive filtering.

Part 2 addresses underdetermined BSS, where there are fewer microphones than source signals. Here, the sparseness of speech sources is very useful; we can utilize time-frequency diversity, where sources are active in different regions of the time-frequency plane.

Part 3 presents monaural BSS where there is only one microphone. Here, we can separate a mixture by using the harmonicity and temporal structure of the sources. We can build a probabilistic framework by assuming a source model, and separate a mixture by maximizing the a posteriori probability of the sources.