Convolutive Residual Echo Power Estimation for Acoustic Echo Reduction

抄録

This paper attempts to estimate the residual echo spectrum that results from the reverberant component beyond a length of a short-time Fourier transform block. A residual echo reduction process suppresses a residual echo signal left after an adaptive filter eliminates an acoustic echo by applying a multiplicative gain calculated from an estimated residual echo power spectrum. However, the power-spectrum estimate in a conventional method reproduces only a fraction of the echo-path impulse response, and the remaining reverberant components are only estimated roughly. Therefore, the conventional method causes large estimation errors of the reverberant components. To address this problem, this paper proposes a finite nonnegative convolution method to estimate the residual echo power spectrum. This method first estimates the power spectra of each segment of echo-impulse response. The method then calculates the residual echo power spectra corresponding to each segment of the residual echo from the power spectra of each segment of the echo-impulse response. Finally, it obtains the residual echo power spectrum that includes the late reverberant components by convoluting their estimated power spectra in a time axis direction. The proposed method estimates all the power spectra in each segment of echo-impulse response by computing cross correlation between the complex spectra of the microphone-input and received speech signals, while shifting present and past frames of the received speech signal in the time axis direction.