The application of frequency‐domain adaptive filters to speech enhancement

Abstract

Several authors have applied the time‐domain least‐mean‐squares adaptive filter to the problem of (voiced) speech enhancement. Generally, these efforts have achieved only limited success due, in part at least, to the nonuniform convergence of the adaptive filter when faced with frequency components of highly disparate spectral power (the so‐called “eigenvalue disparity” problem). This problem is addressed by employing the normalization capacity of the frequency‐domain adaptive filter (FDAF). The first part of this paper deals with the analysis of the FDAF for strictly harmonic signals (this reflects the quasiperiodic nature of voiced speech). It is shown that the behavior of the filler for each weight of the FDAF can be described by a linear transfer function relating the desired input to the output signal. It is further shown that the product of the input power and the feedback component (in each frequency bin) determines the stability and convergence of the filter. Normalization can be achieved by adjusting this product for each weight. Simulations comparing time‐ and frequency‐domain approaches have shown that the FDAF enhancer significantly improves performance.