Wiener solution considering cross-spectral term between echo and near-end speech for acoustic echo reduction

This paper introduces a frequency-domain acoustic echo reduction process based on a new Wiener-filtering method taking into account the cross-spectral term between the acoustic echo and the near-end speech. The conventional echo reduction method based on Wiener filtering estimates the gain based on the assumption that the cross-spectral term of the echo and the near-end speech is zero because the acoustic echo and the near-end speech are statistically uncorrelated. However, this assumption does not always hold true in practice because the gain is estimated in a very short period where the amount of statistical data, which is used to calculate the ensemble averages of the observed signals, is insufficient. As a result, the conventional method occasionally causes the perceptual degradation of sound quality during a double-talk situation; therefore, the performance is still not sufficient. Our goal was to accurately calculate the echo-reduction gain to decrease the speech distortions produced by the echo-reduction process. The proposed method solves a least mean square error of the Wiener-filtering method by taking into account the cross-spectral term between the echo and the near-end speech to obtain a better echo-reduction gain. The performance of this method was demonstrated by objective and subjective results in which speech distortions were decreased.