Sound source separation via computational auditory scene analysis (CASA)-enhanced beamforming.

摘要

In this work, techniques are developed and studied for the extraction of single-source acoustic signals out of multi-source mixtures. Such extracted signals can be used in a variety of applications including: automatic speech recognition, digital hearing aids, teleconferencing, and robot auditory systems. Most previous approaches fall into two categories: computational auditory scene analysis (CASA) and array signal processing.; The approach taken here is to combine these complementary techniques into an integrated one: CASA-enhanced beamforming. This integrated approach has the advantage of combining the array processing location attribute (direction of propagation through a sound-field) with the monaural CASA source attributes (fundamental frequency, on/offset, etc.). The motivation for the CASA-enhanced beamforming approach is the recognition that, by combining the statistically independent location and source attributes, more mixtures can be separated. A mixture that could not be separated by the location attribute alone (for example, if the single-source signals in the mixture have the same location attribute value) may be separated using source attributes, and vice versa.; An alternative to beamforming is binaural CASA. Beamforming is chosen for our integrated approach because it has the following advantages: (1) Since binaural CASA evolved to operate under the constraints of the human auditory system (with only two ears and spectral shaping due to the shape of the human body), it is not clear that it is an ideal method for a computer implementation. Beamforming is more flexible. It allows for any array geometry (number and arrangement of sensors). (2) The beamforming approach is mathematically derived based on a physical model of the acoustic wavefield. So, its processing effect is well-understood. (3) Beamforming operates via an analytic expression. So, its performance can be quantified (as a function of array geometry and the frequency content of the signals in the wavefield).; Experimental results show that CASA-enhanced beamforming extracts wideband signal estimates with higher signal-to-interference ratios (SIR) than monaural CASA, or beamforming alone. That is, it generates wideband signal estimates with the most interference rejection. Regarding intellibility, beamforming produces the lowest spectral distortion. However, CASA-enhanced beamforming's spectral distortion is shown to be comparable to monaural CASA's, and better than binaural CASA's.