Audio Source Separation and Speech Enhancement

11Multichannel Parameter Estimation

Shmulik Markovich‐Golan Walter Kellermann and Sharon Gannot

In this chapter we will explore some widely used structures and estimation procedures for tracking the parameters that are required for constructing the data‐dependent spatial filters for audio signals that are discussed in Chapter 10. As before, the spatial filters are designed to extract a desired source contaminated by background noise, in the case of a single speaker, or by interfering speakers and background noise, in the multiple speakers case.

The spatial filters explored in Chapter 10 (mainly, those referred to as beamformers) assume that certain parameters are available for their computation, namely the relative transfer functions (RTFs) of the speakers, the covariance matrices of the background noise and the speakers, and/or the cross‐covariance between the mixture signals and the desired signal.

The variety of optimization criteria in Chapter 10 leads to different data‐dependent spatial filters, yet most of them rely on similar parameters and therefore a common estimation framework can be derived. In general, estimates of speech presence probability (SPP) are used to govern the estimation of noise and speech spatial covariance matrices. These estimates are then utilized to estimate source RTF vectors. Finally, the data‐dependent spatial filters are designed based on the latter estimates. A high‐level block diagram of the common estimation framework is depicted in Figure 11.1