Preservation of Localization Cues in BSS-Based Noise Reduction: Application in Binaural Hearing Aids

Abstract

For speech applications, blind source separation provides an efficient strategy to enhance the target signal and to reduce the background noise in a noisy environment. Most ICA-based blind source separation (BSS) algorithms are designed under the assumption that the target and interfering signals are spatially located. When the number of interfering signals is small, one of the BSS outputs is expected to provide an excellent estimation of the target signal. Hence, the overall algorithm behaves as an ideal noise-reduction algorithm. However, when the number of interfering signals increases, problem known as the cocktail party effect, or when the background noise is diffusive (i.e., non-point-source noise), this BSS output is no longer a good estimation of the target signal. (Takahashi et al., 2009) showed that in a two-output ICA-based BSS algorithm under these adverse environments, one BSS output includes a mixture of the target signal and residual noise related to the interfering signals, while the other output provides an accurate estimation of the background noise. This particular property validates the experimental results achieved by different post processing strategies to enhance the BSS output associated to the target signal (Noohi & Kahaei, 2010; Parikh et al., 2010; Parikh & Anderson, 2011; Park et al., 2006). These methods are based on Wiener filtering (Kocinski, 2008; Noohi & Kahaei, 2010; Park et al., 2006), spectral subtraction (Kocinski, 2008), least-square (LS) minimization (Parikh et al., 2010), and perceptual post processing (Parikh & Anderson, 2011). All these methods take advantage of a reliable background noise estimator obtained at one of the BSS outputs.