Abstract:
Noise reduction without speech distortion, which is combined with a sub-band acoustic echo canceller in hands-free communications, is proposed. In this method, ambient noise is suppressed by using each sub-band's loss that is adaptively determined based on three conditions: the ratio of the current signal level to the estimated noise level (SNR), previous loss levels, and neighboring sub-band SNRs. The loss level is determined by smoothing out the current loss level and previous ones, where the current loss level is calculated from the sub-band's SNR and neighboring sub-band SNRs. Consequently, this method has no ``musical noise'' due to time-varying loss levels. The noise level used in the calculation of SNR is estimated for each sub-band by maintaining histograms that show the level distribution of the sub-band signals. In addition, this estimated noise level is used simultaneously with the echo signal level to control the affine projection algorithm in a sub-band echo canceller. Using both signal levels, the adaptive algorithm can maintain a stable echo return loss enhancement for small far-end signals. The proposed noise reduction is evaluated in a 32-sub-band echo canceller. It is shown that transmission signals become clearer after reducing echo and ambient noise.