Audio Source Separation and Speech Enhancement

5Spectral Masking and Filtering

Timo Gerkmann and Emmanuel Vincent

In this chapter and the following ones, we consider the case of a single‐channel input signal (). We denote it as and omit the channel index for legibility.

As discussed in Chapter 3, spatial diversity cannot be exploited to separate such a signal due to the difficulty of disambiguating the transfer function from the spectrum of each source. Therefore, single‐channel separation and enhancement must rely on the spectral diversity and exploit properties of the sources such as those listed in Chapter 2. Disregarding phase, one can then separate or enhance the sources using real‐valued filters in the time‐frequency domain known as time‐frequency masks.

In the following, we define the concept of time‐frequency masking in Section 5.1. We introduce different models to derive a mask from the signal statistics in Section 5.2 and modify them in order to improve perceptual quality in Section 5.3. We summarize the main findings and provide links to forthcoming chapters and more advanced topics in Section 5.4.

5.1 Time‐Frequency Masking

5.1.1 Definition and Types of Masks

Following the discussion in Chapter 2, filtering is performed in ...